sync with repo 28.08

README.md

@@ -1,8 +1,35 @@
-ComfyUI
+<div align="center">
-=======
+
-The most powerful and modular stable diffusion GUI and backend.
+# ComfyUI
------------
+**The most powerful and modular stable diffusion GUI and backend.**
+
+
+[![Website][website-shield]][website-url]
+[![Dynamic JSON Badge][discord-shield]][discord-url]
+[![Matrix][matrix-shield]][matrix-url]
+<br>
+[![][github-release-shield]][github-release-link]
+[![][github-release-date-shield]][github-release-link]
+[![][github-downloads-shield]][github-downloads-link]
+[![][github-downloads-latest-shield]][github-downloads-link]
+
+[matrix-shield]: https://img.shields.io/badge/Matrix-000000?style=flat&logo=matrix&logoColor=white
+[matrix-url]: https://app.element.io/#/room/%23comfyui_space%3Amatrix.org
+[website-shield]: https://img.shields.io/badge/ComfyOrg-4285F4?style=flat
+[website-url]: https://www.comfy.org/
+<!-- Workaround to display total user from https://github.com/badges/shields/issues/4500#issuecomment-2060079995 -->
+[discord-shield]: https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fdiscord.com%2Fapi%2Finvites%2Fcomfyorg%3Fwith_counts%3Dtrue&query=%24.approximate_member_count&logo=discord&logoColor=white&label=Discord&color=green&suffix=%20total
+[discord-url]: https://www.comfy.org/discord
+
+[github-release-shield]: https://img.shields.io/github/v/release/comfyanonymous/ComfyUI?style=flat&sort=semver
+[github-release-link]: https://github.com/comfyanonymous/ComfyUI/releases
+[github-release-date-shield]: https://img.shields.io/github/release-date/comfyanonymous/ComfyUI?style=flat
+[github-downloads-shield]: https://img.shields.io/github/downloads/comfyanonymous/ComfyUI/total?style=flat
+[github-downloads-latest-shield]: https://img.shields.io/github/downloads/comfyanonymous/ComfyUI/latest/total?style=flat&label=downloads%40latest
+[github-downloads-link]: https://github.com/comfyanonymous/ComfyUI/releases
+
 ![ComfyUI Screenshot](comfyui_screenshot.png)
+</div>
 
 This ui will let you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. For some workflow examples and see what ComfyUI can do you can check out:
 ### [ComfyUI Examples](https://comfyanonymous.github.io/ComfyUI_examples/)
@@ -48,6 +75,7 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 |------------------------------------|--------------------------------------------------------------------------------------------------------------------|
 | Ctrl + Enter                       | Queue up current graph for generation                                                                              |
 | Ctrl + Shift + Enter               | Queue up current graph as first for generation                                                                     |
+| Ctrl + Alt + Enter                 | Cancel current generation                                                                                          |
 | Ctrl + Z/Ctrl + Y                  | Undo/Redo                                                                                                          |
 | Ctrl + S                           | Save workflow                                                                                                      |
 | Ctrl + O                           | Load workflow                                                                                                      |
@@ -70,6 +98,8 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 | H                                  | Toggle visibility of history                                                                                       |
 | R                                  | Refresh graph                                                                                                      |
 | Double-Click LMB                   | Open node quick search palette                                                                                     |
+| Shift + Drag                       | Move multiple wires at once                                                                                        |
+| Ctrl + Alt + LMB                   | Disconnect all wires from clicked slot                                                                             |
 
 Ctrl can also be replaced with Cmd instead for macOS users
 
@@ -165,20 +195,6 @@ You can install ComfyUI in Apple Mac silicon (M1 or M2) with any recent macOS ve
 
 ```pip install torch-directml``` Then you can launch ComfyUI with: ```python main.py --directml```
 
-### I already have another UI for Stable Diffusion installed do I really have to install all of these dependencies?
-
-You don't. If you have another UI installed and working with its own python venv you can use that venv to run ComfyUI. You can open up your favorite terminal and activate it:
-
-```source path_to_other_sd_gui/venv/bin/activate```
-
-or on Windows:
-
-With Powershell: ```"path_to_other_sd_gui\venv\Scripts\Activate.ps1"```
-
-With cmd.exe: ```"path_to_other_sd_gui\venv\Scripts\activate.bat"```
-
-And then you can use that terminal to run ComfyUI without installing any dependencies. Note that the venv folder might be called something else depending on the SD UI.
-
 # Running
 
 ```python main.py```
@@ -214,7 +230,7 @@ To use a textual inversion concepts/embeddings in a text prompt put them in the
 
 Use ```--preview-method auto``` to enable previews.
 
-The default installation includes a fast latent preview method that's low-resolution. To enable higher-quality previews with [TAESD](https://github.com/madebyollin/taesd), download the [taesd_decoder.pth](https://github.com/madebyollin/taesd/raw/main/taesd_decoder.pth) (for SD1.x and SD2.x) and [taesdxl_decoder.pth](https://github.com/madebyollin/taesd/raw/main/taesdxl_decoder.pth) (for SDXL) models and place them in the `models/vae_approx` folder. Once they're installed, restart ComfyUI to enable high-quality previews.
+The default installation includes a fast latent preview method that's low-resolution. To enable higher-quality previews with [TAESD](https://github.com/madebyollin/taesd), download the [taesd_decoder.pth, taesdxl_decoder.pth, taesd3_decoder.pth and taef1_decoder.pth](https://github.com/madebyollin/taesd/) and place them in the `models/vae_approx` folder. Once they're installed, restart ComfyUI and launch it with `--preview-method taesd` to enable high-quality previews.
 
 ## How to use TLS/SSL?
 Generate a self-signed certificate (not appropriate for shared/production use) and key by running the command: `openssl req -x509 -newkey rsa:4096 -keyout key.pem -out cert.pem -sha256 -days 3650 -nodes -subj "/C=XX/ST=StateName/L=CityName/O=CompanyName/OU=CompanySectionName/CN=CommonNameOrHostname"`
@@ -230,6 +246,47 @@ Use `--tls-keyfile key.pem --tls-certfile cert.pem` to enable TLS/SSL, the app w
 
 See also: [https://www.comfy.org/](https://www.comfy.org/)
 
+## Frontend Development
+
+As of August 15, 2024, we have transitioned to a new frontend, which is now hosted in a separate repository: [ComfyUI Frontend](https://github.com/Comfy-Org/ComfyUI_frontend). This repository now hosts the compiled JS (from TS/Vue) under the `web/` directory.
+
+### Reporting Issues and Requesting Features
+
+For any bugs, issues, or feature requests related to the frontend, please use the [ComfyUI Frontend repository](https://github.com/Comfy-Org/ComfyUI_frontend). This will help us manage and address frontend-specific concerns more efficiently.
+
+### Using the Latest Frontend
+
+The new frontend is now the default for ComfyUI. However, please note:
+
+1. The frontend in the main ComfyUI repository is updated weekly.
+2. Daily releases are available in the separate frontend repository.
+
+To use the most up-to-date frontend version:
+
+1. For the latest daily release, launch ComfyUI with this command line argument:
+
+   ```
+   --front-end-version Comfy-Org/ComfyUI_frontend@latest
+   ```
+
+2. For a specific version, replace `latest` with the desired version number:
+
+   ```
+   --front-end-version Comfy-Org/ComfyUI_frontend@1.2.2
+   ```
+
+This approach allows you to easily switch between the stable weekly release and the cutting-edge daily updates, or even specific versions for testing purposes.
+
+### Accessing the Legacy Frontend
+
+If you need to use the legacy frontend for any reason, you can access it using the following command line argument:
+
+```
+--front-end-version Comfy-Org/ComfyUI_legacy_frontend@latest
+```
+
+This will use a snapshot of the legacy frontend preserved in the [ComfyUI Legacy Frontend repository](https://github.com/Comfy-Org/ComfyUI_legacy_frontend).
+
 # QA
 
 ### Which GPU should I buy for this?

api_server/routes/internal/README.md

@@ -0,0 +1,3 @@
+# ComfyUI Internal Routes
+
+All routes under the `/internal` path are designated for **internal use by ComfyUI only**. These routes are not intended for use by external applications may change at any time without notice.

api_server/routes/internal/internal_routes.py

@@ -0,0 +1,40 @@
+from aiohttp import web
+from typing import Optional
+from folder_paths import models_dir, user_directory, output_directory
+from api_server.services.file_service import FileService
+
+class InternalRoutes:
+    '''
+    The top level web router for internal routes: /internal/*
+    The endpoints here should NOT be depended upon. It is for ComfyUI frontend use only.
+    Check README.md for more information.
+    
+    '''
+    def __init__(self):
+        self.routes: web.RouteTableDef = web.RouteTableDef()
+        self._app: Optional[web.Application] = None
+        self.file_service = FileService({
+            "models": models_dir,
+            "user": user_directory,
+            "output": output_directory
+        })
+
+    def setup_routes(self):
+        @self.routes.get('/files')
+        async def list_files(request):
+            directory_key = request.query.get('directory', '')
+            try:
+                file_list = self.file_service.list_files(directory_key)
+                return web.json_response({"files": file_list})
+            except ValueError as e:
+                return web.json_response({"error": str(e)}, status=400)
+            except Exception as e:
+                return web.json_response({"error": str(e)}, status=500)
+
+
+    def get_app(self):
+        if self._app is None:
+            self._app = web.Application()
+            self.setup_routes()
+            self._app.add_routes(self.routes)
+        return self._app

api_server/services/file_service.py

@@ -0,0 +1,13 @@
+from typing import Dict, List, Optional
+from api_server.utils.file_operations import FileSystemOperations, FileSystemItem
+
+class FileService:
+    def __init__(self, allowed_directories: Dict[str, str], file_system_ops: Optional[FileSystemOperations] = None):
+        self.allowed_directories: Dict[str, str] = allowed_directories
+        self.file_system_ops: FileSystemOperations = file_system_ops or FileSystemOperations()
+
+    def list_files(self, directory_key: str) -> List[FileSystemItem]:
+        if directory_key not in self.allowed_directories:
+            raise ValueError("Invalid directory key")
+        directory_path: str = self.allowed_directories[directory_key]
+        return self.file_system_ops.walk_directory(directory_path)

api_server/utils/file_operations.py

@@ -0,0 +1,42 @@
+import os
+from typing import List, Union, TypedDict, Literal
+from typing_extensions import TypeGuard
+class FileInfo(TypedDict):
+    name: str
+    path: str
+    type: Literal["file"]
+    size: int
+
+class DirectoryInfo(TypedDict):
+    name: str
+    path: str
+    type: Literal["directory"]
+
+FileSystemItem = Union[FileInfo, DirectoryInfo]
+
+def is_file_info(item: FileSystemItem) -> TypeGuard[FileInfo]:
+    return item["type"] == "file"
+
+class FileSystemOperations:
+    @staticmethod
+    def walk_directory(directory: str) -> List[FileSystemItem]:
+        file_list: List[FileSystemItem] = []
+        for root, dirs, files in os.walk(directory):
+            for name in files:
+                file_path = os.path.join(root, name)
+                relative_path = os.path.relpath(file_path, directory)
+                file_list.append({
+                    "name": name,
+                    "path": relative_path,
+                    "type": "file",
+                    "size": os.path.getsize(file_path)
+                })
+            for name in dirs:
+                dir_path = os.path.join(root, name)
+                relative_path = os.path.relpath(dir_path, directory)
+                file_list.append({
+                    "name": name,
+                    "path": relative_path,
+                    "type": "directory"
+                })
+        return file_list

app/frontend_management.py

@@ -8,7 +8,7 @@ import zipfile
 from dataclasses import dataclass
 from functools import cached_property
 from pathlib import Path
-from typing import TypedDict
+from typing import TypedDict, Optional
 
 import requests
 from typing_extensions import NotRequired
@@ -132,12 +132,13 @@ class FrontendManager:
         return match_result.group(1), match_result.group(2), match_result.group(3)
 
     @classmethod
-    def init_frontend_unsafe(cls, version_string: str) -> str:
+    def init_frontend_unsafe(cls, version_string: str, provider: Optional[FrontEndProvider] = None) -> str:
         """
         Initializes the frontend for the specified version.
 
         Args:
             version_string (str): The version string.
+            provider (FrontEndProvider, optional): The provider to use. Defaults to None.
 
         Returns:
             str: The path to the initialized frontend.
@@ -150,7 +151,7 @@ class FrontendManager:
             return cls.DEFAULT_FRONTEND_PATH
 
         repo_owner, repo_name, version = cls.parse_version_string(version_string)
-        provider = FrontEndProvider(repo_owner, repo_name)
+        provider = provider or FrontEndProvider(repo_owner, repo_name)
         release = provider.get_release(version)
 
         semantic_version = release["tag_name"].lstrip("v")
@@ -158,15 +159,21 @@ class FrontendManager:
             Path(cls.CUSTOM_FRONTENDS_ROOT) / provider.folder_name / semantic_version
         )
         if not os.path.exists(web_root):
-            os.makedirs(web_root, exist_ok=True)
+            try:
-            logging.info(
+                os.makedirs(web_root, exist_ok=True)
-                "Downloading frontend(%s) version(%s) to (%s)",
+                logging.info(
-                provider.folder_name,
+                    "Downloading frontend(%s) version(%s) to (%s)",
-                semantic_version,
+                    provider.folder_name,
-                web_root,
+                    semantic_version,
-            )
+                    web_root,
-            logging.debug(release)
+                )
-            download_release_asset_zip(release, destination_path=web_root)
+                logging.debug(release)
+                download_release_asset_zip(release, destination_path=web_root)
+            finally:
+                # Clean up the directory if it is empty, i.e. the download failed
+                if not os.listdir(web_root):
+                    os.rmdir(web_root)
+
         return web_root
 
     @classmethod

comfy/cli_args.py

@@ -92,6 +92,10 @@ class LatentPreviewMethod(enum.Enum):
 
 parser.add_argument("--preview-method", type=LatentPreviewMethod, default=LatentPreviewMethod.NoPreviews, help="Default preview method for sampler nodes.", action=EnumAction)
 
+cache_group = parser.add_mutually_exclusive_group()
+cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
+cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
+
 attn_group = parser.add_mutually_exclusive_group()
 attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.")
 attn_group.add_argument("--use-quad-cross-attention", action="store_true", help="Use the sub-quadratic cross attention optimization . Ignored when xformers is used.")
@@ -112,10 +116,14 @@ vram_group.add_argument("--lowvram", action="store_true", help="Split the unet i
 vram_group.add_argument("--novram", action="store_true", help="When lowvram isn't enough.")
 vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for everything (slow).")
 
+parser.add_argument("--reserve-vram", type=float, default=None, help="Set the amount of vram in GB you want to reserve for use by your OS/other software. By default some amount is reverved depending on your OS.")
+
+
 parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'sha1', 'sha256', 'sha512'], default='sha256', help="Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256.")
 
 parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")
 parser.add_argument("--deterministic", action="store_true", help="Make pytorch use slower deterministic algorithms when it can. Note that this might not make images deterministic in all cases.")
+parser.add_argument("--fast", action="store_true", help="Enable some untested and potentially quality deteriorating optimizations.")
 
 parser.add_argument("--dont-print-server", action="store_true", help="Don't print server output.")
 parser.add_argument("--quick-test-for-ci", action="store_true", help="Quick test for CI.")

comfy/clip_model.py

@@ -88,10 +88,11 @@ class CLIPTextModel_(torch.nn.Module):
         heads = config_dict["num_attention_heads"]
         intermediate_size = config_dict["intermediate_size"]
         intermediate_activation = config_dict["hidden_act"]
+        num_positions = config_dict["max_position_embeddings"]
         self.eos_token_id = config_dict["eos_token_id"]
 
         super().__init__()
-        self.embeddings = CLIPEmbeddings(embed_dim, dtype=dtype, device=device, operations=operations)
+        self.embeddings = CLIPEmbeddings(embed_dim, num_positions=num_positions, dtype=dtype, device=device, operations=operations)
         self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
         self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
 
@@ -123,7 +124,6 @@ class CLIPTextModel(torch.nn.Module):
         self.text_model = CLIPTextModel_(config_dict, dtype, device, operations)
         embed_dim = config_dict["hidden_size"]
         self.text_projection = operations.Linear(embed_dim, embed_dim, bias=False, dtype=dtype, device=device)
-        self.text_projection.weight.copy_(torch.eye(embed_dim))
         self.dtype = dtype
 
     def get_input_embeddings(self):

comfy/controlnet.py

@@ -1,4 +1,24 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
+
 import torch
+from enum import Enum
 import math
 import os
 import logging
@@ -13,6 +33,8 @@ import comfy.cldm.cldm
 import comfy.t2i_adapter.adapter
 import comfy.ldm.cascade.controlnet
 import comfy.cldm.mmdit
+import comfy.ldm.hydit.controlnet
+import comfy.ldm.flux.controlnet_xlabs
 
 
 def broadcast_image_to(tensor, target_batch_size, batched_number):
@@ -33,6 +55,10 @@ def broadcast_image_to(tensor, target_batch_size, batched_number):
     else:
         return torch.cat([tensor] * batched_number, dim=0)
 
+class StrengthType(Enum):
+    CONSTANT = 1
+    LINEAR_UP = 2
+
 class ControlBase:
     def __init__(self, device=None):
         self.cond_hint_original = None
@@ -51,6 +77,8 @@ class ControlBase:
             device = comfy.model_management.get_torch_device()
         self.device = device
         self.previous_controlnet = None
+        self.extra_conds = []
+        self.strength_type = StrengthType.CONSTANT
 
     def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(0.0, 1.0), vae=None):
         self.cond_hint_original = cond_hint
@@ -93,6 +121,8 @@ class ControlBase:
         c.latent_format = self.latent_format
         c.extra_args = self.extra_args.copy()
         c.vae = self.vae
+        c.extra_conds = self.extra_conds.copy()
+        c.strength_type = self.strength_type
 
     def inference_memory_requirements(self, dtype):
         if self.previous_controlnet is not None:
@@ -113,7 +143,10 @@ class ControlBase:
 
                     if x not in applied_to: #memory saving strategy, allow shared tensors and only apply strength to shared tensors once
                         applied_to.add(x)
-                        x *= self.strength
+                        if self.strength_type == StrengthType.CONSTANT:
+                            x *= self.strength
+                        elif self.strength_type == StrengthType.LINEAR_UP:
+                            x *= (self.strength ** float(len(control_output) - i))
 
                     if x.dtype != output_dtype:
                         x = x.to(output_dtype)
@@ -142,7 +175,7 @@ class ControlBase:
 
 
 class ControlNet(ControlBase):
-    def __init__(self, control_model=None, global_average_pooling=False, compression_ratio=8, latent_format=None, device=None, load_device=None, manual_cast_dtype=None):
+    def __init__(self, control_model=None, global_average_pooling=False, compression_ratio=8, latent_format=None, device=None, load_device=None, manual_cast_dtype=None, extra_conds=["y"], strength_type=StrengthType.CONSTANT):
         super().__init__(device)
         self.control_model = control_model
         self.load_device = load_device
@@ -154,6 +187,8 @@ class ControlNet(ControlBase):
         self.model_sampling_current = None
         self.manual_cast_dtype = manual_cast_dtype
         self.latent_format = latent_format
+        self.extra_conds += extra_conds
+        self.strength_type = strength_type
 
     def get_control(self, x_noisy, t, cond, batched_number):
         control_prev = None
@@ -191,13 +226,16 @@ class ControlNet(ControlBase):
             self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
 
         context = cond.get('crossattn_controlnet', cond['c_crossattn'])
-        y = cond.get('y', None)
+        extra = self.extra_args.copy()
-        if y is not None:
+        for c in self.extra_conds:
-            y = y.to(dtype)
+            temp = cond.get(c, None)
+            if temp is not None:
+                extra[c] = temp.to(dtype)
+
         timestep = self.model_sampling_current.timestep(t)
         x_noisy = self.model_sampling_current.calculate_input(t, x_noisy)
 
-        control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.float(), context=context.to(dtype), y=y, **self.extra_args)
+        control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.to(dtype), context=context.to(dtype), **extra)
         return self.control_merge(control, control_prev, output_dtype)
 
     def copy(self):
@@ -286,6 +324,7 @@ class ControlLora(ControlNet):
         ControlBase.__init__(self, device)
         self.control_weights = control_weights
         self.global_average_pooling = global_average_pooling
+        self.extra_conds += ["y"]
 
     def pre_run(self, model, percent_to_timestep_function):
         super().pre_run(model, percent_to_timestep_function)
@@ -338,12 +377,8 @@ class ControlLora(ControlNet):
     def inference_memory_requirements(self, dtype):
         return comfy.utils.calculate_parameters(self.control_weights) * comfy.model_management.dtype_size(dtype) + ControlBase.inference_memory_requirements(self, dtype)
 
-def load_controlnet_mmdit(sd):
+def controlnet_config(sd):
-    new_sd = comfy.model_detection.convert_diffusers_mmdit(sd, "")
+    model_config = comfy.model_detection.model_config_from_unet(sd, "", True)
-    model_config = comfy.model_detection.model_config_from_unet(new_sd, "", True)
-    num_blocks = comfy.model_detection.count_blocks(new_sd, 'joint_blocks.{}.')
-    for k in sd:
-        new_sd[k] = sd[k]
 
     supported_inference_dtypes = model_config.supported_inference_dtypes
 
@@ -356,14 +391,28 @@ def load_controlnet_mmdit(sd):
     else:
         operations = comfy.ops.disable_weight_init
 
-    control_model = comfy.cldm.mmdit.ControlNet(num_blocks=num_blocks, operations=operations, device=load_device, dtype=unet_dtype, **controlnet_config)
+    offload_device = comfy.model_management.unet_offload_device()
-    missing, unexpected = control_model.load_state_dict(new_sd, strict=False)
+    return model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device
+
+def controlnet_load_state_dict(control_model, sd):
+    missing, unexpected = control_model.load_state_dict(sd, strict=False)
 
     if len(missing) > 0:
         logging.warning("missing controlnet keys: {}".format(missing))
 
     if len(unexpected) > 0:
         logging.debug("unexpected controlnet keys: {}".format(unexpected))
+    return control_model
+
+def load_controlnet_mmdit(sd):
+    new_sd = comfy.model_detection.convert_diffusers_mmdit(sd, "")
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(new_sd)
+    num_blocks = comfy.model_detection.count_blocks(new_sd, 'joint_blocks.{}.')
+    for k in sd:
+        new_sd[k] = sd[k]
+
+    control_model = comfy.cldm.mmdit.ControlNet(num_blocks=num_blocks, operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
+    control_model = controlnet_load_state_dict(control_model, new_sd)
 
     latent_format = comfy.latent_formats.SD3()
     latent_format.shift_factor = 0 #SD3 controlnet weirdness
@@ -371,8 +420,31 @@ def load_controlnet_mmdit(sd):
     return control
 
 
+def load_controlnet_hunyuandit(controlnet_data):
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(controlnet_data)
+
+    control_model = comfy.ldm.hydit.controlnet.HunYuanControlNet(operations=operations, device=offload_device, dtype=unet_dtype)
+    control_model = controlnet_load_state_dict(control_model, controlnet_data)
+
+    latent_format = comfy.latent_formats.SDXL()
+    extra_conds = ['text_embedding_mask', 'encoder_hidden_states_t5', 'text_embedding_mask_t5', 'image_meta_size', 'style', 'cos_cis_img', 'sin_cis_img']
+    control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds, strength_type=StrengthType.CONSTANT)
+    return control
+
+def load_controlnet_flux_xlabs(sd):
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(sd)
+    control_model = comfy.ldm.flux.controlnet_xlabs.ControlNetFlux(operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
+    control_model = controlnet_load_state_dict(control_model, sd)
+    extra_conds = ['y', 'guidance']
+    control = ControlNet(control_model, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
+    return control
+
+
 def load_controlnet(ckpt_path, model=None):
     controlnet_data = comfy.utils.load_torch_file(ckpt_path, safe_load=True)
+    if 'after_proj_list.18.bias' in controlnet_data.keys(): #Hunyuan DiT
+        return load_controlnet_hunyuandit(controlnet_data)
+
     if "lora_controlnet" in controlnet_data:
         return ControlLora(controlnet_data)
 
@@ -430,7 +502,10 @@ def load_controlnet(ckpt_path, model=None):
             logging.warning("leftover keys: {}".format(leftover_keys))
         controlnet_data = new_sd
     elif "controlnet_blocks.0.weight" in controlnet_data: #SD3 diffusers format
-        return load_controlnet_mmdit(controlnet_data)
+        if "double_blocks.0.img_attn.norm.key_norm.scale" in controlnet_data:
+            return load_controlnet_flux_xlabs(controlnet_data)
+        else:
+            return load_controlnet_mmdit(controlnet_data)
 
     pth_key = 'control_model.zero_convs.0.0.weight'
     pth = False
@@ -462,6 +537,7 @@ def load_controlnet(ckpt_path, model=None):
     if manual_cast_dtype is not None:
         controlnet_config["operations"] = comfy.ops.manual_cast
     controlnet_config["dtype"] = unet_dtype
+    controlnet_config["device"] = comfy.model_management.unet_offload_device()
     controlnet_config.pop("out_channels")
     controlnet_config["hint_channels"] = controlnet_data["{}input_hint_block.0.weight".format(prefix)].shape[1]
     control_model = comfy.cldm.cldm.ControlNet(**controlnet_config)

comfy/diffusers_load.py

@@ -22,7 +22,7 @@ def load_diffusers(model_path, output_vae=True, output_clip=True, embedding_dire
     if text_encoder2_path is not None:
         text_encoder_paths.append(text_encoder2_path)
 
-    unet = comfy.sd.load_unet(unet_path)
+    unet = comfy.sd.load_diffusion_model(unet_path)
 
     clip = None
     if output_clip:

comfy/float.py

@@ -0,0 +1,62 @@
+import torch
+import math
+
+def calc_mantissa(abs_x, exponent, normal_mask, MANTISSA_BITS, EXPONENT_BIAS, generator=None):
+    mantissa_scaled = torch.where(
+        normal_mask,
+        (abs_x / (2.0 ** (exponent - EXPONENT_BIAS)) - 1.0) * (2**MANTISSA_BITS),
+        (abs_x / (2.0 ** (-EXPONENT_BIAS + 1 - MANTISSA_BITS)))
+    )
+
+    mantissa_scaled += torch.rand(mantissa_scaled.size(), dtype=mantissa_scaled.dtype, layout=mantissa_scaled.layout, device=mantissa_scaled.device, generator=generator)
+    return mantissa_scaled.floor() / (2**MANTISSA_BITS)
+
+#Not 100% sure about this
+def manual_stochastic_round_to_float8(x, dtype, generator=None):
+    if dtype == torch.float8_e4m3fn:
+        EXPONENT_BITS, MANTISSA_BITS, EXPONENT_BIAS = 4, 3, 7
+    elif dtype == torch.float8_e5m2:
+        EXPONENT_BITS, MANTISSA_BITS, EXPONENT_BIAS = 5, 2, 15
+    else:
+        raise ValueError("Unsupported dtype")
+
+    x = x.half()
+    sign = torch.sign(x)
+    abs_x = x.abs()
+    sign = torch.where(abs_x == 0, 0, sign)
+
+    # Combine exponent calculation and clamping
+    exponent = torch.clamp(
+        torch.floor(torch.log2(abs_x)) + EXPONENT_BIAS,
+        0, 2**EXPONENT_BITS - 1
+    )
+
+    # Combine mantissa calculation and rounding
+    normal_mask = ~(exponent == 0)
+
+    abs_x[:] = calc_mantissa(abs_x, exponent, normal_mask, MANTISSA_BITS, EXPONENT_BIAS, generator=generator)
+
+    sign *= torch.where(
+        normal_mask,
+        (2.0 ** (exponent - EXPONENT_BIAS)) * (1.0 + abs_x),
+        (2.0 ** (-EXPONENT_BIAS + 1)) * abs_x
+    )
+    del abs_x
+
+    return sign.to(dtype=dtype)
+
+
+
+def stochastic_rounding(value, dtype, seed=0):
+    if dtype == torch.float32:
+        return value.to(dtype=torch.float32)
+    if dtype == torch.float16:
+        return value.to(dtype=torch.float16)
+    if dtype == torch.bfloat16:
+        return value.to(dtype=torch.bfloat16)
+    if dtype == torch.float8_e4m3fn or dtype == torch.float8_e5m2:
+        generator = torch.Generator(device=value.device)
+        generator.manual_seed(seed)
+        return manual_stochastic_round_to_float8(value, dtype, generator=generator)
+
+    return value.to(dtype=dtype)

comfy/k_diffusion/sampling.py

@@ -9,6 +9,7 @@ from tqdm.auto import trange, tqdm
 from . import utils
 from . import deis
 import comfy.model_patcher
+import comfy.model_sampling
 
 def append_zero(x):
     return torch.cat([x, x.new_zeros([1])])
@@ -509,6 +510,9 @@ def sample_dpm_adaptive(model, x, sigma_min, sigma_max, extra_args=None, callbac
 
 @torch.no_grad()
 def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    if isinstance(model.inner_model.inner_model.model_sampling, comfy.model_sampling.CONST):
+        return sample_dpmpp_2s_ancestral_RF(model, x, sigmas, extra_args, callback, disable, eta, s_noise, noise_sampler)
+
     """Ancestral sampling with DPM-Solver++(2S) second-order steps."""
     extra_args = {} if extra_args is None else extra_args
     noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
@@ -541,6 +545,55 @@ def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None,
     return x
 
 
+@torch.no_grad()
+def sample_dpmpp_2s_ancestral_RF(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    """Ancestral sampling with DPM-Solver++(2S) second-order steps."""
+    extra_args = {} if extra_args is None else extra_args
+    noise_sampler = default_noise_sampler(x) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+    sigma_fn = lambda lbda: (lbda.exp() + 1) ** -1
+    lambda_fn = lambda sigma: ((1-sigma)/sigma).log()
+
+    # logged_x = x.unsqueeze(0)
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        downstep_ratio = 1 + (sigmas[i+1]/sigmas[i] - 1) * eta
+        sigma_down = sigmas[i+1] * downstep_ratio
+        alpha_ip1 = 1 - sigmas[i+1]
+        alpha_down = 1 - sigma_down
+        renoise_coeff = (sigmas[i+1]**2 - sigma_down**2*alpha_ip1**2/alpha_down**2)**0.5
+        # sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            # Euler method
+            d = to_d(x, sigmas[i], denoised)
+            dt = sigma_down - sigmas[i]
+            x = x + d * dt
+        else:
+            # DPM-Solver++(2S)
+            if sigmas[i] == 1.0:
+                sigma_s = 0.9999
+            else:
+                t_i, t_down = lambda_fn(sigmas[i]), lambda_fn(sigma_down)
+                r = 1 / 2
+                h = t_down - t_i
+                s = t_i + r * h
+                sigma_s = sigma_fn(s)
+            # sigma_s = sigmas[i+1]
+            sigma_s_i_ratio = sigma_s / sigmas[i]
+            u = sigma_s_i_ratio * x + (1 - sigma_s_i_ratio) * denoised
+            D_i = model(u, sigma_s * s_in, **extra_args)
+            sigma_down_i_ratio = sigma_down / sigmas[i]
+            x = sigma_down_i_ratio * x + (1 - sigma_down_i_ratio) * D_i
+            # print("sigma_i", sigmas[i], "sigma_ip1", sigmas[i+1],"sigma_down", sigma_down, "sigma_down_i_ratio", sigma_down_i_ratio, "sigma_s_i_ratio", sigma_s_i_ratio, "renoise_coeff", renoise_coeff)
+        # Noise addition
+        if sigmas[i + 1] > 0 and eta > 0:
+            x = (alpha_ip1/alpha_down) * x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * renoise_coeff
+        # logged_x = torch.cat((logged_x, x.unsqueeze(0)), dim=0)
+    return x
+
 @torch.no_grad()
 def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
     """DPM-Solver++ (stochastic)."""

comfy/latent_formats.py

@@ -141,6 +141,7 @@ class StableAudio1(LatentFormat):
     latent_channels = 64
 
 class Flux(SD3):
+    latent_channels = 16
     def __init__(self):
         self.scale_factor = 0.3611
         self.shift_factor = 0.1159
@@ -162,6 +163,7 @@ class Flux(SD3):
             [-0.0005, -0.0530, -0.0020],
             [-0.1273, -0.0932, -0.0680]
         ]
+        self.taesd_decoder_name = "taef1_decoder"
 
     def process_in(self, latent):
         return (latent - self.shift_factor) * self.scale_factor

comfy/ldm/aura/mmdit.py

@@ -9,6 +9,7 @@ import torch.nn.functional as F
 
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.ops
+import comfy.ldm.common_dit
 
 def modulate(x, shift, scale):
     return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
@@ -407,10 +408,7 @@ class MMDiT(nn.Module):
 
     def patchify(self, x):
         B, C, H, W = x.size()
-        pad_h = (self.patch_size - H % self.patch_size) % self.patch_size
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
-        pad_w = (self.patch_size - W % self.patch_size) % self.patch_size
-
-        x = torch.nn.functional.pad(x, (0, pad_w, 0, pad_h), mode='circular')
         x = x.view(
             B,
             C,

comfy/ldm/common_dit.py

@@ -0,0 +1,8 @@
+import torch
+
+def pad_to_patch_size(img, patch_size=(2, 2), padding_mode="circular"):
+    if padding_mode == "circular" and torch.jit.is_tracing() or torch.jit.is_scripting():
+        padding_mode = "reflect"
+    pad_h = (patch_size[0] - img.shape[-2] % patch_size[0]) % patch_size[0]
+    pad_w = (patch_size[1] - img.shape[-1] % patch_size[1]) % patch_size[1]
+    return torch.nn.functional.pad(img, (0, pad_w, 0, pad_h), mode=padding_mode)

comfy/ldm/flux/controlnet_xlabs.py

@@ -0,0 +1,104 @@
+#Original code can be found on: https://github.com/XLabs-AI/x-flux/blob/main/src/flux/controlnet.py
+
+import torch
+from torch import Tensor, nn
+from einops import rearrange, repeat
+
+from .layers import (DoubleStreamBlock, EmbedND, LastLayer,
+                                 MLPEmbedder, SingleStreamBlock,
+                                 timestep_embedding)
+
+from .model import Flux
+import comfy.ldm.common_dit
+
+
+class ControlNetFlux(Flux):
+    def __init__(self, image_model=None, dtype=None, device=None, operations=None, **kwargs):
+        super().__init__(final_layer=False, dtype=dtype, device=device, operations=operations, **kwargs)
+
+        # add ControlNet blocks
+        self.controlnet_blocks = nn.ModuleList([])
+        for _ in range(self.params.depth):
+            controlnet_block = operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
+            # controlnet_block = zero_module(controlnet_block)
+            self.controlnet_blocks.append(controlnet_block)
+        self.pos_embed_input = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
+        self.gradient_checkpointing = False
+        self.input_hint_block = nn.Sequential(
+            operations.Conv2d(3, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device)
+        )
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        controlnet_cond: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        y: Tensor,
+        guidance: Tensor = None,
+    ) -> Tensor:
+        if img.ndim != 3 or txt.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        # running on sequences img
+        img = self.img_in(img)
+        controlnet_cond = self.input_hint_block(controlnet_cond)
+        controlnet_cond = rearrange(controlnet_cond, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+        controlnet_cond = self.pos_embed_input(controlnet_cond)
+        img = img + controlnet_cond
+        vec = self.time_in(timestep_embedding(timesteps, 256))
+        if self.params.guidance_embed:
+            vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
+        vec = vec + self.vector_in(y)
+        txt = self.txt_in(txt)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        block_res_samples = ()
+
+        for block in self.double_blocks:
+            img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
+            block_res_samples = block_res_samples + (img,)
+
+        controlnet_block_res_samples = ()
+        for block_res_sample, controlnet_block in zip(block_res_samples, self.controlnet_blocks):
+            block_res_sample = controlnet_block(block_res_sample)
+            controlnet_block_res_samples = controlnet_block_res_samples + (block_res_sample,)
+
+        return {"input": (controlnet_block_res_samples * 10)[:19]}
+
+    def forward(self, x, timesteps, context, y, guidance=None, hint=None, **kwargs):
+        hint = hint * 2.0 - 1.0
+
+        bs, c, h, w = x.shape
+        patch_size = 2
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
+
+        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
+
+        h_len = ((h + (patch_size // 2)) // patch_size)
+        w_len = ((w + (patch_size // 2)) // patch_size)
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[..., 1] = img_ids[..., 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype)[:, None]
+        img_ids[..., 2] = img_ids[..., 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype)[None, :]
+        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        return self.forward_orig(img, img_ids, hint, context, txt_ids, timesteps, y, guidance)

comfy/ldm/flux/layers.py

@@ -2,12 +2,12 @@ import math
 from dataclasses import dataclass
 
 import torch
-from einops import rearrange
 from torch import Tensor, nn
 
 from .math import attention, rope
 import comfy.ops
 
+
 class EmbedND(nn.Module):
     def __init__(self, dim: int, theta: int, axes_dim: list):
         super().__init__()
@@ -36,9 +36,7 @@ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 10
     """
     t = time_factor * t
     half = dim // 2
-    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=t.device) / half)
-        t.device
-    )
 
     args = t[:, None].float() * freqs[None]
     embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
@@ -48,7 +46,6 @@ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 10
         embedding = embedding.to(t)
     return embedding
 
-
 class MLPEmbedder(nn.Module):
     def __init__(self, in_dim: int, hidden_dim: int, dtype=None, device=None, operations=None):
         super().__init__()
@@ -66,10 +63,8 @@ class RMSNorm(torch.nn.Module):
         self.scale = nn.Parameter(torch.empty((dim), dtype=dtype, device=device))
 
     def forward(self, x: Tensor):
-        x_dtype = x.dtype
-        x = x.float()
         rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + 1e-6)
-        return (x * rrms).to(dtype=x_dtype) * comfy.ops.cast_to(self.scale, dtype=x_dtype, device=x.device)
+        return (x * rrms) * comfy.ops.cast_to(self.scale, dtype=x.dtype, device=x.device)
 
 
 class QKNorm(torch.nn.Module):
@@ -94,14 +89,6 @@ class SelfAttention(nn.Module):
         self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
         self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
 
-    def forward(self, x: Tensor, pe: Tensor) -> Tensor:
-        qkv = self.qkv(x)
-        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
-        q, k = self.norm(q, k, v)
-        x = attention(q, k, v, pe=pe)
-        x = self.proj(x)
-        return x
-
 
 @dataclass
 class ModulationOut:
@@ -163,22 +150,21 @@ class DoubleStreamBlock(nn.Module):
         img_modulated = self.img_norm1(img)
         img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
         img_qkv = self.img_attn.qkv(img_modulated)
-        img_q, img_k, img_v = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
         img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
 
         # prepare txt for attention
         txt_modulated = self.txt_norm1(txt)
         txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
         txt_qkv = self.txt_attn.qkv(txt_modulated)
-        txt_q, txt_k, txt_v = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
         txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
 
         # run actual attention
-        q = torch.cat((txt_q, img_q), dim=2)
+        attn = attention(torch.cat((txt_q, img_q), dim=2),
-        k = torch.cat((txt_k, img_k), dim=2)
+                         torch.cat((txt_k, img_k), dim=2),
-        v = torch.cat((txt_v, img_v), dim=2)
+                         torch.cat((txt_v, img_v), dim=2), pe=pe)
 
-        attn = attention(q, k, v, pe=pe)
         txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
 
         # calculate the img bloks
@@ -186,8 +172,12 @@ class DoubleStreamBlock(nn.Module):
         img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift)
 
         # calculate the txt bloks
-        txt = txt + txt_mod1.gate * self.txt_attn.proj(txt_attn)
+        txt += txt_mod1.gate * self.txt_attn.proj(txt_attn)
-        txt = txt + txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
+        txt += txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
+
+        if txt.dtype == torch.float16:
+            txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504)
+
         return img, txt
 
 
@@ -232,14 +222,17 @@ class SingleStreamBlock(nn.Module):
         x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
         qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
 
-        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
         q, k = self.norm(q, k, v)
 
         # compute attention
         attn = attention(q, k, v, pe=pe)
         # compute activation in mlp stream, cat again and run second linear layer
         output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
-        return x + mod.gate * output
+        x += mod.gate * output
+        if x.dtype == torch.float16:
+            x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
+        return x
 
 
 class LastLayer(nn.Module):

comfy/ldm/flux/math.py

@@ -14,7 +14,7 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor) -> Tensor:
 
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
-    if comfy.model_management.is_device_mps(pos.device):
+    if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu():
         device = torch.device("cpu")
     else:
         device = pos.device

comfy/ldm/flux/model.py

@@ -15,6 +15,7 @@ from .layers import (
 )
 
 from einops import rearrange, repeat
+import comfy.ldm.common_dit
 
 @dataclass
 class FluxParams:
@@ -37,12 +38,12 @@ class Flux(nn.Module):
     Transformer model for flow matching on sequences.
     """
 
-    def __init__(self, image_model=None, dtype=None, device=None, operations=None, **kwargs):
+    def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
         super().__init__()
         self.dtype = dtype
         params = FluxParams(**kwargs)
         self.params = params
-        self.in_channels = params.in_channels
+        self.in_channels = params.in_channels * 2 * 2
         self.out_channels = self.in_channels
         if params.hidden_size % params.num_heads != 0:
             raise ValueError(
@@ -82,7 +83,8 @@ class Flux(nn.Module):
             ]
         )
 
-        self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
+        if final_layer:
+            self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
 
     def forward_orig(
         self,
@@ -93,6 +95,7 @@ class Flux(nn.Module):
         timesteps: Tensor,
         y: Tensor,
         guidance: Tensor = None,
+        control=None,
     ) -> Tensor:
         if img.ndim != 3 or txt.ndim != 3:
             raise ValueError("Input img and txt tensors must have 3 dimensions.")
@@ -111,24 +114,37 @@ class Flux(nn.Module):
         ids = torch.cat((txt_ids, img_ids), dim=1)
         pe = self.pe_embedder(ids)
 
-        for block in self.double_blocks:
+        for i, block in enumerate(self.double_blocks):
             img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
 
+            if control is not None: # Controlnet
+                control_i = control.get("input")
+                if i < len(control_i):
+                    add = control_i[i]
+                    if add is not None:
+                        img += add
+
         img = torch.cat((txt, img), 1)
-        for block in self.single_blocks:
+
+        for i, block in enumerate(self.single_blocks):
             img = block(img, vec=vec, pe=pe)
+
+            if control is not None: # Controlnet
+                control_o = control.get("output")
+                if i < len(control_o):
+                    add = control_o[i]
+                    if add is not None:
+                        img[:, txt.shape[1] :, ...] += add
+
         img = img[:, txt.shape[1] :, ...]
 
         img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
         return img
 
-    def forward(self, x, timestep, context, y, guidance, **kwargs):
+    def forward(self, x, timestep, context, y, guidance, control=None, **kwargs):
         bs, c, h, w = x.shape
         patch_size = 2
-        pad_h = (patch_size - h % 2) % patch_size
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
-        pad_w = (patch_size - w % 2) % patch_size
-
-        x = torch.nn.functional.pad(x, (0, pad_w, 0, pad_h), mode='circular')
 
         img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
 
@@ -140,5 +156,5 @@ class Flux(nn.Module):
         img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
 
         txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
-        out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance)
+        out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control)
         return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h,:w]

comfy/ldm/hydit/attn_layers.py

@@ -47,7 +47,7 @@ def reshape_for_broadcast(freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]], x
 
 
 def rotate_half(x):
-    x_real, x_imag = x.float().reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+    x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
     return torch.stack([-x_imag, x_real], dim=-1).flatten(3)
 
 
@@ -78,10 +78,9 @@ def apply_rotary_emb(
     xk_out = None
     if isinstance(freqs_cis, tuple):
         cos, sin = reshape_for_broadcast(freqs_cis, xq, head_first)    # [S, D]
-        cos, sin = cos.to(xq.device), sin.to(xq.device)
+        xq_out = (xq * cos + rotate_half(xq) * sin)
-        xq_out = (xq.float() * cos + rotate_half(xq.float()) * sin).type_as(xq)
         if xk is not None:
-            xk_out = (xk.float() * cos + rotate_half(xk.float()) * sin).type_as(xk)
+            xk_out = (xk * cos + rotate_half(xk) * sin)
     else:
         xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))  # [B, S, H, D//2]
         freqs_cis = reshape_for_broadcast(freqs_cis, xq_, head_first).to(xq.device)   # [S, D//2] --> [1, S, 1, D//2]

comfy/ldm/hydit/controlnet.py

@@ -0,0 +1,321 @@
+from typing import Any, Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from torch.utils import checkpoint
+
+from comfy.ldm.modules.diffusionmodules.mmdit import (
+    Mlp,
+    TimestepEmbedder,
+    PatchEmbed,
+    RMSNorm,
+)
+from comfy.ldm.modules.diffusionmodules.util import timestep_embedding
+from .poolers import AttentionPool
+
+import comfy.latent_formats
+from .models import HunYuanDiTBlock, calc_rope
+
+from .posemb_layers import get_2d_rotary_pos_embed, get_fill_resize_and_crop
+
+
+class HunYuanControlNet(nn.Module):
+    """
+    HunYuanDiT: Diffusion model with a Transformer backbone.
+
+    Inherit ModelMixin and ConfigMixin to be compatible with the sampler StableDiffusionPipeline of diffusers.
+
+    Inherit PeftAdapterMixin to be compatible with the PEFT training pipeline.
+
+    Parameters
+    ----------
+    args: argparse.Namespace
+        The arguments parsed by argparse.
+    input_size: tuple
+        The size of the input image.
+    patch_size: int
+        The size of the patch.
+    in_channels: int
+        The number of input channels.
+    hidden_size: int
+        The hidden size of the transformer backbone.
+    depth: int
+        The number of transformer blocks.
+    num_heads: int
+        The number of attention heads.
+    mlp_ratio: float
+        The ratio of the hidden size of the MLP in the transformer block.
+    log_fn: callable
+        The logging function.
+    """
+
+    def __init__(
+        self,
+        input_size: tuple = 128,
+        patch_size: int = 2,
+        in_channels: int = 4,
+        hidden_size: int = 1408,
+        depth: int = 40,
+        num_heads: int = 16,
+        mlp_ratio: float = 4.3637,
+        text_states_dim=1024,
+        text_states_dim_t5=2048,
+        text_len=77,
+        text_len_t5=256,
+        qk_norm=True,  # See http://arxiv.org/abs/2302.05442 for details.
+        size_cond=False,
+        use_style_cond=False,
+        learn_sigma=True,
+        norm="layer",
+        log_fn: callable = print,
+        attn_precision=None,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs,
+    ):
+        super().__init__()
+        self.log_fn = log_fn
+        self.depth = depth
+        self.learn_sigma = learn_sigma
+        self.in_channels = in_channels
+        self.out_channels = in_channels * 2 if learn_sigma else in_channels
+        self.patch_size = patch_size
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.text_states_dim = text_states_dim
+        self.text_states_dim_t5 = text_states_dim_t5
+        self.text_len = text_len
+        self.text_len_t5 = text_len_t5
+        self.size_cond = size_cond
+        self.use_style_cond = use_style_cond
+        self.norm = norm
+        self.dtype = dtype
+        self.latent_format = comfy.latent_formats.SDXL
+
+        self.mlp_t5 = nn.Sequential(
+            nn.Linear(
+                self.text_states_dim_t5,
+                self.text_states_dim_t5 * 4,
+                bias=True,
+                dtype=dtype,
+                device=device,
+            ),
+            nn.SiLU(),
+            nn.Linear(
+                self.text_states_dim_t5 * 4,
+                self.text_states_dim,
+                bias=True,
+                dtype=dtype,
+                device=device,
+            ),
+        )
+        # learnable replace
+        self.text_embedding_padding = nn.Parameter(
+            torch.randn(
+                self.text_len + self.text_len_t5,
+                self.text_states_dim,
+                dtype=dtype,
+                device=device,
+            )
+        )
+
+        # Attention pooling
+        pooler_out_dim = 1024
+        self.pooler = AttentionPool(
+            self.text_len_t5,
+            self.text_states_dim_t5,
+            num_heads=8,
+            output_dim=pooler_out_dim,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        # Dimension of the extra input vectors
+        self.extra_in_dim = pooler_out_dim
+
+        if self.size_cond:
+            # Image size and crop size conditions
+            self.extra_in_dim += 6 * 256
+
+        if self.use_style_cond:
+            # Here we use a default learned embedder layer for future extension.
+            self.style_embedder = nn.Embedding(
+                1, hidden_size, dtype=dtype, device=device
+            )
+            self.extra_in_dim += hidden_size
+
+        # Text embedding for `add`
+        self.x_embedder = PatchEmbed(
+            input_size,
+            patch_size,
+            in_channels,
+            hidden_size,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.t_embedder = TimestepEmbedder(
+            hidden_size, dtype=dtype, device=device, operations=operations
+        )
+        self.extra_embedder = nn.Sequential(
+            operations.Linear(
+                self.extra_in_dim, hidden_size * 4, dtype=dtype, device=device
+            ),
+            nn.SiLU(),
+            operations.Linear(
+                hidden_size * 4, hidden_size, bias=True, dtype=dtype, device=device
+            ),
+        )
+
+        # Image embedding
+        num_patches = self.x_embedder.num_patches
+
+        # HUnYuanDiT Blocks
+        self.blocks = nn.ModuleList(
+            [
+                HunYuanDiTBlock(
+                    hidden_size=hidden_size,
+                    c_emb_size=hidden_size,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    text_states_dim=self.text_states_dim,
+                    qk_norm=qk_norm,
+                    norm_type=self.norm,
+                    skip=False,
+                    attn_precision=attn_precision,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations,
+                )
+                for _ in range(19)
+            ]
+        )
+
+        # Input zero linear for the first block
+        self.before_proj = operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
+
+
+        # Output zero linear for the every block
+        self.after_proj_list = nn.ModuleList(
+            [
+
+                    operations.Linear(
+                        self.hidden_size, self.hidden_size, dtype=dtype, device=device
+                    )
+                for _ in range(len(self.blocks))
+            ]
+        )
+
+    def forward(
+        self,
+        x,
+        hint,
+        timesteps,
+        context,#encoder_hidden_states=None,
+        text_embedding_mask=None,
+        encoder_hidden_states_t5=None,
+        text_embedding_mask_t5=None,
+        image_meta_size=None,
+        style=None,
+        return_dict=False,
+        **kwarg,
+    ):
+        """
+        Forward pass of the encoder.
+
+        Parameters
+        ----------
+        x: torch.Tensor
+            (B, D, H, W)
+        t: torch.Tensor
+            (B)
+        encoder_hidden_states: torch.Tensor
+            CLIP text embedding, (B, L_clip, D)
+        text_embedding_mask: torch.Tensor
+            CLIP text embedding mask, (B, L_clip)
+        encoder_hidden_states_t5: torch.Tensor
+            T5 text embedding, (B, L_t5, D)
+        text_embedding_mask_t5: torch.Tensor
+            T5 text embedding mask, (B, L_t5)
+        image_meta_size: torch.Tensor
+            (B, 6)
+        style: torch.Tensor
+            (B)
+        cos_cis_img: torch.Tensor
+        sin_cis_img: torch.Tensor
+        return_dict: bool
+            Whether to return a dictionary.
+        """
+        condition = hint
+        if condition.shape[0] == 1:
+            condition = torch.repeat_interleave(condition, x.shape[0], dim=0)
+
+        text_states = context  # 2,77,1024
+        text_states_t5 = encoder_hidden_states_t5  # 2,256,2048
+        text_states_mask = text_embedding_mask.bool()  # 2,77
+        text_states_t5_mask = text_embedding_mask_t5.bool()  # 2,256
+        b_t5, l_t5, c_t5 = text_states_t5.shape
+        text_states_t5 = self.mlp_t5(text_states_t5.view(-1, c_t5)).view(b_t5, l_t5, -1)
+
+        padding = comfy.ops.cast_to_input(self.text_embedding_padding, text_states)
+
+        text_states[:, -self.text_len :] = torch.where(
+            text_states_mask[:, -self.text_len :].unsqueeze(2),
+            text_states[:, -self.text_len :],
+            padding[: self.text_len],
+        )
+        text_states_t5[:, -self.text_len_t5 :] = torch.where(
+            text_states_t5_mask[:, -self.text_len_t5 :].unsqueeze(2),
+            text_states_t5[:, -self.text_len_t5 :],
+            padding[self.text_len :],
+        )
+
+        text_states = torch.cat([text_states, text_states_t5], dim=1)  # 2,205，1024
+
+        # _, _, oh, ow = x.shape
+        # th, tw = oh // self.patch_size, ow // self.patch_size
+
+        # Get image RoPE embedding according to `reso`lution.
+        freqs_cis_img = calc_rope(
+            x, self.patch_size, self.hidden_size // self.num_heads
+        )  # (cos_cis_img, sin_cis_img)
+
+        # ========================= Build time and image embedding =========================
+        t = self.t_embedder(timesteps, dtype=self.dtype)
+        x = self.x_embedder(x)
+
+        # ========================= Concatenate all extra vectors =========================
+        # Build text tokens with pooling
+        extra_vec = self.pooler(encoder_hidden_states_t5)
+
+        # Build image meta size tokens if applicable
+        # if image_meta_size is not None:
+        #     image_meta_size = timestep_embedding(image_meta_size.view(-1), 256)   # [B * 6, 256]
+        #     if image_meta_size.dtype != self.dtype:
+        #         image_meta_size = image_meta_size.half()
+        #     image_meta_size = image_meta_size.view(-1, 6 * 256)
+        #     extra_vec = torch.cat([extra_vec, image_meta_size], dim=1)  # [B, D + 6 * 256]
+
+        # Build style tokens
+        if style is not None:
+            style_embedding = self.style_embedder(style)
+            extra_vec = torch.cat([extra_vec, style_embedding], dim=1)
+
+        # Concatenate all extra vectors
+        c = t + self.extra_embedder(extra_vec)  # [B, D]
+
+        # ========================= Deal with Condition =========================
+        condition = self.x_embedder(condition)
+
+        # ========================= Forward pass through HunYuanDiT blocks =========================
+        controls = []
+        x = x + self.before_proj(condition)  # add condition
+        for layer, block in enumerate(self.blocks):
+            x = block(x, c, text_states, freqs_cis_img)
+            controls.append(self.after_proj_list[layer](x))  # zero linear for output
+
+        return {"output": controls}

comfy/ldm/hydit/models.py

@@ -21,6 +21,7 @@ def calc_rope(x, patch_size, head_size):
     sub_args = [start, stop, (th, tw)]
     # head_size = HUNYUAN_DIT_CONFIG['DiT-g/2']['hidden_size'] // HUNYUAN_DIT_CONFIG['DiT-g/2']['num_heads']
     rope = get_2d_rotary_pos_embed(head_size, *sub_args)
+    rope = (rope[0].to(x), rope[1].to(x))
     return rope
 
 
@@ -91,6 +92,8 @@ class HunYuanDiTBlock(nn.Module):
         # Long Skip Connection
         if self.skip_linear is not None:
             cat = torch.cat([x, skip], dim=-1)
+            if cat.dtype != x.dtype:
+                cat = cat.to(x.dtype)
             cat = self.skip_norm(cat)
             x = self.skip_linear(cat)
 
@@ -362,6 +365,8 @@ class HunYuanDiT(nn.Module):
         c = t + self.extra_embedder(extra_vec)  # [B, D]
 
         controls = None
+        if control:
+            controls = control.get("output", None)
         # ========================= Forward pass through HunYuanDiT blocks =========================
         skips = []
         for layer, block in enumerate(self.blocks):

comfy/ldm/modules/attention.py

@@ -358,7 +358,7 @@ def attention_xformers(q, k, v, heads, mask=None, attn_precision=None, skip_resh
             disabled_xformers = True
 
     if disabled_xformers:
-        return attention_pytorch(q, k, v, heads, mask)
+        return attention_pytorch(q, k, v, heads, mask, skip_reshape=skip_reshape)
 
     if skip_reshape:
          q, k, v = map(

comfy/ldm/modules/diffusionmodules/mmdit.py

@@ -9,6 +9,7 @@ from .. import attention
 from einops import rearrange, repeat
 from .util import timestep_embedding
 import comfy.ops
+import comfy.ldm.common_dit
 
 def default(x, y):
     if x is not None:
@@ -111,9 +112,7 @@ class PatchEmbed(nn.Module):
         #             f"Input width ({W}) should be divisible by patch size ({self.patch_size[1]})."
         #         )
         if self.dynamic_img_pad:
-            pad_h = (self.patch_size[0] - H % self.patch_size[0]) % self.patch_size[0]
+            x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size, padding_mode=self.padding_mode)
-            pad_w = (self.patch_size[1] - W % self.patch_size[1]) % self.patch_size[1]
-            x = torch.nn.functional.pad(x, (0, pad_w, 0, pad_h), mode=self.padding_mode)
         x = self.proj(x)
         if self.flatten:
             x = x.flatten(2).transpose(1, 2)  # NCHW -> NLC

comfy/lora.py

@@ -1,5 +1,27 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
+from __future__ import annotations
 import comfy.utils
+import comfy.model_management
+import comfy.model_base
 import logging
+import torch
 
 LORA_CLIP_MAP = {
     "mlp.fc1": "mlp_fc1",
@@ -218,11 +240,17 @@ def model_lora_keys_clip(model, key_map={}):
                     lora_key = "lora_prior_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #cascade lora: TODO put lora key prefix in the model config
                     key_map[lora_key] = k
 
-    for k in sdk: #OneTrainer SD3 lora
+    for k in sdk:
-        if k.startswith("t5xxl.transformer.") and k.endswith(".weight"):
+        if k.endswith(".weight"):
-            l_key = k[len("t5xxl.transformer."):-len(".weight")]
+            if k.startswith("t5xxl.transformer."):#OneTrainer SD3 lora
-            lora_key = "lora_te3_{}".format(l_key.replace(".", "_"))
+                l_key = k[len("t5xxl.transformer."):-len(".weight")]
-            key_map[lora_key] = k
+                lora_key = "lora_te3_{}".format(l_key.replace(".", "_"))
+                key_map[lora_key] = k
+            elif k.startswith("hydit_clip.transformer.bert."): #HunyuanDiT Lora
+                l_key = k[len("hydit_clip.transformer.bert."):-len(".weight")]
+                lora_key = "lora_te1_{}".format(l_key.replace(".", "_"))
+                key_map[lora_key] = k
+
 
     k = "clip_g.transformer.text_projection.weight"
     if k in sdk:
@@ -245,6 +273,7 @@ def model_lora_keys_unet(model, key_map={}):
             key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
             key_map["lora_unet_{}".format(key_lora)] = k
             key_map["lora_prior_unet_{}".format(key_lora)] = k #cascade lora: TODO put lora key prefix in the model config
+            key_map["{}".format(k[:-len(".weight")])] = k #generic lora format without any weird key names
 
     diffusers_keys = comfy.utils.unet_to_diffusers(model.model_config.unet_config)
     for k in diffusers_keys:
@@ -288,4 +317,240 @@ def model_lora_keys_unet(model, key_map={}):
                 key_lora = k[len("diffusion_model."):-len(".weight")]
                 key_map["base_model.model.{}".format(key_lora)] = k #official hunyuan lora format
 
+    if isinstance(model, comfy.model_base.Flux): #Diffusers lora Flux
+        diffusers_keys = comfy.utils.flux_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.")
+        for k in diffusers_keys:
+            if k.endswith(".weight"):
+                to = diffusers_keys[k]
+                key_map["transformer.{}".format(k[:-len(".weight")])] = to #simpletrainer and probably regular diffusers flux lora format
+                key_map["lycoris_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #simpletrainer lycoris
+
     return key_map
+
+
+def weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype):
+    dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, intermediate_dtype)
+    lora_diff *= alpha
+    weight_calc = weight + lora_diff.type(weight.dtype)
+    weight_norm = (
+        weight_calc.transpose(0, 1)
+        .reshape(weight_calc.shape[1], -1)
+        .norm(dim=1, keepdim=True)
+        .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1))
+        .transpose(0, 1)
+    )
+
+    weight_calc *= (dora_scale / weight_norm).type(weight.dtype)
+    if strength != 1.0:
+        weight_calc -= weight
+        weight += strength * (weight_calc)
+    else:
+        weight[:] = weight_calc
+    return weight
+
+def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Tensor:
+    """
+    Pad a tensor to a new shape with zeros.
+
+    Args:
+        tensor (torch.Tensor): The original tensor to be padded.
+        new_shape (List[int]): The desired shape of the padded tensor.
+
+    Returns:
+        torch.Tensor: A new tensor padded with zeros to the specified shape.
+
+    Note:
+        If the new shape is smaller than the original tensor in any dimension,
+        the original tensor will be truncated in that dimension.
+    """
+    if any([new_shape[i] < tensor.shape[i] for i in range(len(new_shape))]):
+        raise ValueError("The new shape must be larger than the original tensor in all dimensions")
+
+    if len(new_shape) != len(tensor.shape):
+        raise ValueError("The new shape must have the same number of dimensions as the original tensor")
+
+    # Create a new tensor filled with zeros
+    padded_tensor = torch.zeros(new_shape, dtype=tensor.dtype, device=tensor.device)
+
+    # Create slicing tuples for both tensors
+    orig_slices = tuple(slice(0, dim) for dim in tensor.shape)
+    new_slices = tuple(slice(0, dim) for dim in tensor.shape)
+
+    # Copy the original tensor into the new tensor
+    padded_tensor[new_slices] = tensor[orig_slices]
+
+    return padded_tensor
+
+def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32):
+    for p in patches:
+        strength = p[0]
+        v = p[1]
+        strength_model = p[2]
+        offset = p[3]
+        function = p[4]
+        if function is None:
+            function = lambda a: a
+
+        old_weight = None
+        if offset is not None:
+            old_weight = weight
+            weight = weight.narrow(offset[0], offset[1], offset[2])
+
+        if strength_model != 1.0:
+            weight *= strength_model
+
+        if isinstance(v, list):
+            v = (calculate_weight(v[1:], v[0].clone(), key, intermediate_dtype=intermediate_dtype), )
+
+        if len(v) == 1:
+            patch_type = "diff"
+        elif len(v) == 2:
+            patch_type = v[0]
+            v = v[1]
+
+        if patch_type == "diff":
+            diff: torch.Tensor = v[0]
+            # An extra flag to pad the weight if the diff's shape is larger than the weight
+            do_pad_weight = len(v) > 1 and v[1]['pad_weight']
+            if do_pad_weight and diff.shape != weight.shape:
+                logging.info("Pad weight {} from {} to shape: {}".format(key, weight.shape, diff.shape))
+                weight = pad_tensor_to_shape(weight, diff.shape)
+
+            if strength != 0.0:
+                if diff.shape != weight.shape:
+                    logging.warning("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED {} != {}".format(key, diff.shape, weight.shape))
+                else:
+                    weight += function(strength * comfy.model_management.cast_to_device(diff, weight.device, weight.dtype))
+        elif patch_type == "lora": #lora/locon
+            mat1 = comfy.model_management.cast_to_device(v[0], weight.device, intermediate_dtype)
+            mat2 = comfy.model_management.cast_to_device(v[1], weight.device, intermediate_dtype)
+            dora_scale = v[4]
+            if v[2] is not None:
+                alpha = v[2] / mat2.shape[0]
+            else:
+                alpha = 1.0
+
+            if v[3] is not None:
+                #locon mid weights, hopefully the math is fine because I didn't properly test it
+                mat3 = comfy.model_management.cast_to_device(v[3], weight.device, intermediate_dtype)
+                final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
+                mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1)
+            try:
+                lora_diff = torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)).reshape(weight.shape)
+                if dora_scale is not None:
+                    weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
+                else:
+                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+            except Exception as e:
+                logging.error("ERROR {} {} {}".format(patch_type, key, e))
+        elif patch_type == "lokr":
+            w1 = v[0]
+            w2 = v[1]
+            w1_a = v[3]
+            w1_b = v[4]
+            w2_a = v[5]
+            w2_b = v[6]
+            t2 = v[7]
+            dora_scale = v[8]
+            dim = None
+
+            if w1 is None:
+                dim = w1_b.shape[0]
+                w1 = torch.mm(comfy.model_management.cast_to_device(w1_a, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w1_b, weight.device, intermediate_dtype))
+            else:
+                w1 = comfy.model_management.cast_to_device(w1, weight.device, intermediate_dtype)
+
+            if w2 is None:
+                dim = w2_b.shape[0]
+                if t2 is None:
+                    w2 = torch.mm(comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype))
+                else:
+                    w2 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                        comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
+                                        comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype),
+                                        comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype))
+            else:
+                w2 = comfy.model_management.cast_to_device(w2, weight.device, intermediate_dtype)
+
+            if len(w2.shape) == 4:
+                w1 = w1.unsqueeze(2).unsqueeze(2)
+            if v[2] is not None and dim is not None:
+                alpha = v[2] / dim
+            else:
+                alpha = 1.0
+
+            try:
+                lora_diff = torch.kron(w1, w2).reshape(weight.shape)
+                if dora_scale is not None:
+                    weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
+                else:
+                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+            except Exception as e:
+                logging.error("ERROR {} {} {}".format(patch_type, key, e))
+        elif patch_type == "loha":
+            w1a = v[0]
+            w1b = v[1]
+            if v[2] is not None:
+                alpha = v[2] / w1b.shape[0]
+            else:
+                alpha = 1.0
+
+            w2a = v[3]
+            w2b = v[4]
+            dora_scale = v[7]
+            if v[5] is not None: #cp decomposition
+                t1 = v[5]
+                t2 = v[6]
+                m1 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                    comfy.model_management.cast_to_device(t1, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype))
+
+                m2 = torch.einsum('i j k l, j r, i p -> p r k l',
+                                    comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype),
+                                    comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype))
+            else:
+                m1 = torch.mm(comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype))
+                m2 = torch.mm(comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype),
+                                comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype))
+
+            try:
+                lora_diff = (m1 * m2).reshape(weight.shape)
+                if dora_scale is not None:
+                    weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
+                else:
+                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+            except Exception as e:
+                logging.error("ERROR {} {} {}".format(patch_type, key, e))
+        elif patch_type == "glora":
+            if v[4] is not None:
+                alpha = v[4] / v[0].shape[0]
+            else:
+                alpha = 1.0
+
+            dora_scale = v[5]
+
+            a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, intermediate_dtype)
+            a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, intermediate_dtype)
+            b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, intermediate_dtype)
+            b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, intermediate_dtype)
+
+            try:
+                lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1), a2), a1)).reshape(weight.shape)
+                if dora_scale is not None:
+                    weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype))
+                else:
+                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
+            except Exception as e:
+                logging.error("ERROR {} {} {}".format(patch_type, key, e))
+        else:
+            logging.warning("patch type not recognized {} {}".format(patch_type, key))
+
+        if old_weight is not None:
+            weight = old_weight
+
+    return weight

comfy/model_base.py

@@ -1,3 +1,21 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
 import torch
 import logging
 from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep
@@ -74,16 +92,18 @@ class BaseModel(torch.nn.Module):
         self.latent_format = model_config.latent_format
         self.model_config = model_config
         self.manual_cast_dtype = model_config.manual_cast_dtype
+        self.device = device
 
         if not unet_config.get("disable_unet_model_creation", False):
-            if self.manual_cast_dtype is not None:
+            if model_config.custom_operations is None:
-                operations = comfy.ops.manual_cast
+                operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype)
             else:
-                operations = comfy.ops.disable_weight_init
+                operations = model_config.custom_operations
             self.diffusion_model = unet_model(**unet_config, device=device, operations=operations)
             if comfy.model_management.force_channels_last():
                 self.diffusion_model.to(memory_format=torch.channels_last)
                 logging.debug("using channels last mode for diffusion model")
+            logging.info("model weight dtype {}, manual cast: {}".format(self.get_dtype(), self.manual_cast_dtype))
         self.model_type = model_type
         self.model_sampling = model_sampling(model_config, model_type)
 
@@ -94,6 +114,7 @@ class BaseModel(torch.nn.Module):
         self.concat_keys = ()
         logging.info("model_type {}".format(model_type.name))
         logging.debug("adm {}".format(self.adm_channels))
+        self.memory_usage_factor = model_config.memory_usage_factor
 
     def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
         sigma = t
@@ -252,11 +273,11 @@ class BaseModel(torch.nn.Module):
                 dtype = self.manual_cast_dtype
             #TODO: this needs to be tweaked
             area = input_shape[0] * math.prod(input_shape[2:])
-            return (area * comfy.model_management.dtype_size(dtype) / 50) * (1024 * 1024)
+            return (area * comfy.model_management.dtype_size(dtype) * 0.01 * self.memory_usage_factor) * (1024 * 1024)
         else:
             #TODO: this formula might be too aggressive since I tweaked the sub-quad and split algorithms to use less memory.
             area = input_shape[0] * math.prod(input_shape[2:])
-            return (area * 0.3) * (1024 * 1024)
+            return (area * 0.15 * self.memory_usage_factor) * (1024 * 1024)
 
 
 def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge=0.0, seed=None):
@@ -354,6 +375,7 @@ class SDXL(BaseModel):
         flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0).repeat(clip_pooled.shape[0], 1)
         return torch.cat((clip_pooled.to(flat.device), flat), dim=1)
 
+
 class SVD_img2vid(BaseModel):
     def __init__(self, model_config, model_type=ModelType.V_PREDICTION_EDM, device=None):
         super().__init__(model_config, model_type, device=device)
@@ -594,17 +616,6 @@ class SD3(BaseModel):
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         return out
 
-    def memory_required(self, input_shape):
-        if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
-            dtype = self.get_dtype()
-            if self.manual_cast_dtype is not None:
-                dtype = self.manual_cast_dtype
-            #TODO: this probably needs to be tweaked
-            area = input_shape[0] * input_shape[2] * input_shape[3]
-            return (area * comfy.model_management.dtype_size(dtype) * 0.012) * (1024 * 1024)
-        else:
-            area = input_shape[0] * input_shape[2] * input_shape[3]
-            return (area * 0.3) * (1024 * 1024)
 
 class AuraFlow(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
@@ -702,15 +713,3 @@ class Flux(BaseModel):
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([kwargs.get("guidance", 3.5)]))
         return out
-
-    def memory_required(self, input_shape):
-        if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
-            dtype = self.get_dtype()
-            if self.manual_cast_dtype is not None:
-                dtype = self.manual_cast_dtype
-            #TODO: this probably needs to be tweaked
-            area = input_shape[0] * input_shape[2] * input_shape[3]
-            return (area * comfy.model_management.dtype_size(dtype) * 0.020) * (1024 * 1024)
-        else:
-            area = input_shape[0] * input_shape[2] * input_shape[3]
-            return (area * 0.3) * (1024 * 1024)

comfy/model_detection.py

@@ -131,14 +131,14 @@ def detect_unet_config(state_dict, key_prefix):
     if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys: #Flux
         dit_config = {}
         dit_config["image_model"] = "flux"
-        dit_config["in_channels"] = 64
+        dit_config["in_channels"] = 16
         dit_config["vec_in_dim"] = 768
         dit_config["context_in_dim"] = 4096
         dit_config["hidden_size"] = 3072
         dit_config["mlp_ratio"] = 4.0
         dit_config["num_heads"] = 24
-        dit_config["depth"] = 19
+        dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
-        dit_config["depth_single_blocks"] = 38
+        dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
         dit_config["axes_dim"] = [16, 56, 56]
         dit_config["theta"] = 10000
         dit_config["qkv_bias"] = True
@@ -472,9 +472,15 @@ def unet_config_from_diffusers_unet(state_dict, dtype=None):
             'transformer_depth': [0, 1, 1], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': -2, 'use_linear_in_transformer': False,
             'context_dim': 768, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 1, 1, 1, 1],
             'use_temporal_attention': False, 'use_temporal_resblock': False}
+    
+    SD15_diffusers_inpaint = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'adm_in_channels': None,
+            'dtype': dtype, 'in_channels': 9, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
+            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': False, 'context_dim': 768, 'num_heads': 8,
+            'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}  
 
 
-    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p]
+    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p, SD15_diffusers_inpaint]
 
     for unet_config in supported_models:
         matches = True
@@ -495,7 +501,12 @@ def model_config_from_diffusers_unet(state_dict):
 def convert_diffusers_mmdit(state_dict, output_prefix=""):
     out_sd = {}
 
-    if 'transformer_blocks.0.attn.add_q_proj.weight' in state_dict: #SD3
+    if 'transformer_blocks.0.attn.norm_added_k.weight' in state_dict: #Flux
+        depth = count_blocks(state_dict, 'transformer_blocks.{}.')
+        depth_single_blocks = count_blocks(state_dict, 'single_transformer_blocks.{}.')
+        hidden_size = state_dict["x_embedder.bias"].shape[0]
+        sd_map = comfy.utils.flux_to_diffusers({"depth": depth, "depth_single_blocks": depth_single_blocks, "hidden_size": hidden_size}, output_prefix=output_prefix)
+    elif 'transformer_blocks.0.attn.add_q_proj.weight' in state_dict: #SD3
         num_blocks = count_blocks(state_dict, 'transformer_blocks.{}.')
         depth = state_dict["pos_embed.proj.weight"].shape[0] // 64
         sd_map = comfy.utils.mmdit_to_diffusers({"depth": depth, "num_blocks": num_blocks}, output_prefix=output_prefix)
@@ -521,7 +532,12 @@ def convert_diffusers_mmdit(state_dict, output_prefix=""):
                     old_weight = out_sd.get(t[0], None)
                     if old_weight is None:
                         old_weight = torch.empty_like(weight)
-                        old_weight = old_weight.repeat([3] + [1] * (len(old_weight.shape) - 1))
+                    if old_weight.shape[offset[0]] < offset[1] + offset[2]:
+                        exp = list(weight.shape)
+                        exp[offset[0]] = offset[1] + offset[2]
+                        new = torch.empty(exp, device=weight.device, dtype=weight.dtype)
+                        new[:old_weight.shape[0]] = old_weight
+                        old_weight = new
 
                     w = old_weight.narrow(offset[0], offset[1], offset[2])
                 else:

comfy/model_management.py

@@ -1,3 +1,21 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
 import psutil
 import logging
 from enum import Enum
@@ -26,9 +44,14 @@ cpu_state = CPUState.GPU
 
 total_vram = 0
 
-lowvram_available = True
 xpu_available = False
+try:
+    torch_version = torch.version.__version__
+    xpu_available = (int(torch_version[0]) < 2 or (int(torch_version[0]) == 2 and int(torch_version[2]) <= 4)) and torch.xpu.is_available()
+except:
+    pass
 
+lowvram_available = True
 if args.deterministic:
     logging.info("Using deterministic algorithms for pytorch")
     torch.use_deterministic_algorithms(True, warn_only=True)
@@ -48,10 +71,10 @@ if args.directml is not None:
 
 try:
     import intel_extension_for_pytorch as ipex
-    if torch.xpu.is_available():
+    _ = torch.xpu.device_count()
-        xpu_available = True
+    xpu_available = torch.xpu.is_available()
 except:
-    pass
+    xpu_available = xpu_available or (hasattr(torch, "xpu") and torch.xpu.is_available())
 
 try:
     if torch.backends.mps.is_available():
@@ -171,7 +194,6 @@ VAE_DTYPES = [torch.float32]
 
 try:
     if is_nvidia():
-        torch_version = torch.version.__version__
         if int(torch_version[0]) >= 2:
             if ENABLE_PYTORCH_ATTENTION == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
                 ENABLE_PYTORCH_ATTENTION = True
@@ -273,9 +295,12 @@ class LoadedModel:
     def model_memory(self):
         return self.model.model_size()
 
+    def model_offloaded_memory(self):
+        return self.model.model_size() - self.model.loaded_size()
+
     def model_memory_required(self, device):
-        if device == self.model.current_device:
+        if device == self.model.current_loaded_device():
-            return 0
+            return self.model_offloaded_memory()
         else:
             return self.model_memory()
 
@@ -287,39 +312,77 @@ class LoadedModel:
 
         load_weights = not self.weights_loaded
 
-        try:
+        if self.model.loaded_size() > 0:
-            if lowvram_model_memory > 0 and load_weights:
+            use_more_vram = lowvram_model_memory
-                self.real_model = self.model.patch_model_lowvram(device_to=patch_model_to, lowvram_model_memory=lowvram_model_memory, force_patch_weights=force_patch_weights)
+            if use_more_vram == 0:
-            else:
+                use_more_vram = 1e32
-                self.real_model = self.model.patch_model(device_to=patch_model_to, patch_weights=load_weights)
+            self.model_use_more_vram(use_more_vram)
-        except Exception as e:
+        else:
-            self.model.unpatch_model(self.model.offload_device)
+            try:
-            self.model_unload()
+                self.real_model = self.model.patch_model(device_to=patch_model_to, lowvram_model_memory=lowvram_model_memory, load_weights=load_weights, force_patch_weights=force_patch_weights)
-            raise e
+            except Exception as e:
+                self.model.unpatch_model(self.model.offload_device)
+                self.model_unload()
+                raise e
 
-        if is_intel_xpu() and not args.disable_ipex_optimize:
+        if is_intel_xpu() and not args.disable_ipex_optimize and self.real_model is not None:
-            self.real_model = ipex.optimize(self.real_model.eval(), graph_mode=True, concat_linear=True)
+            with torch.no_grad():
+                self.real_model = ipex.optimize(self.real_model.eval(), inplace=True, graph_mode=True, concat_linear=True)
 
         self.weights_loaded = True
         return self.real_model
 
     def should_reload_model(self, force_patch_weights=False):
-        if force_patch_weights and self.model.lowvram_patch_counter > 0:
+        if force_patch_weights and self.model.lowvram_patch_counter() > 0:
             return True
         return False
 
-    def model_unload(self, unpatch_weights=True):
+    def model_unload(self, memory_to_free=None, unpatch_weights=True):
+        if memory_to_free is not None:
+            if memory_to_free < self.model.loaded_size():
+                freed = self.model.partially_unload(self.model.offload_device, memory_to_free)
+                if freed >= memory_to_free:
+                    return False
         self.model.unpatch_model(self.model.offload_device, unpatch_weights=unpatch_weights)
         self.model.model_patches_to(self.model.offload_device)
         self.weights_loaded = self.weights_loaded and not unpatch_weights
         self.real_model = None
+        return True
+
+    def model_use_more_vram(self, extra_memory):
+        return self.model.partially_load(self.device, extra_memory)
 
     def __eq__(self, other):
         return self.model is other.model
 
+def use_more_memory(extra_memory, loaded_models, device):
+    for m in loaded_models:
+        if m.device == device:
+            extra_memory -= m.model_use_more_vram(extra_memory)
+            if extra_memory <= 0:
+                break
+
+def offloaded_memory(loaded_models, device):
+    offloaded_mem = 0
+    for m in loaded_models:
+        if m.device == device:
+            offloaded_mem += m.model_offloaded_memory()
+    return offloaded_mem
+
 def minimum_inference_memory():
     return (1024 * 1024 * 1024) * 1.2
 
+EXTRA_RESERVED_VRAM = 200 * 1024 * 1024
+if any(platform.win32_ver()):
+    EXTRA_RESERVED_VRAM = 500 * 1024 * 1024 #Windows is higher because of the shared vram issue
+
+if args.reserve_vram is not None:
+    EXTRA_RESERVED_VRAM = args.reserve_vram * 1024 * 1024 * 1024
+    logging.debug("Reserving {}MB vram for other applications.".format(EXTRA_RESERVED_VRAM / (1024 * 1024)))
+
+def extra_reserved_memory():
+    return EXTRA_RESERVED_VRAM
+
 def unload_model_clones(model, unload_weights_only=True, force_unload=True):
     to_unload = []
     for i in range(len(current_loaded_models)):
@@ -342,6 +405,8 @@ def unload_model_clones(model, unload_weights_only=True, force_unload=True):
     if not force_unload:
         if unload_weights_only and unload_weight == False:
             return None
+    else:
+        unload_weight = True
 
     for i in to_unload:
         logging.debug("unload clone {} {}".format(i, unload_weight))
@@ -352,6 +417,7 @@ def unload_model_clones(model, unload_weights_only=True, force_unload=True):
 def free_memory(memory_required, device, keep_loaded=[]):
     unloaded_model = []
     can_unload = []
+    unloaded_models = []
 
     for i in range(len(current_loaded_models) -1, -1, -1):
         shift_model = current_loaded_models[i]
@@ -362,14 +428,18 @@ def free_memory(memory_required, device, keep_loaded=[]):
 
     for x in sorted(can_unload):
         i = x[-1]
+        memory_to_free = None
         if not DISABLE_SMART_MEMORY:
-            if get_free_memory(device) > memory_required:
+            free_mem = get_free_memory(device)
+            if free_mem > memory_required:
                 break
-        current_loaded_models[i].model_unload()
+            memory_to_free = memory_required - free_mem
-        unloaded_model.append(i)
+        logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
+        if current_loaded_models[i].model_unload(memory_to_free):
+            unloaded_model.append(i)
 
     for i in sorted(unloaded_model, reverse=True):
-        current_loaded_models.pop(i)
+        unloaded_models.append(current_loaded_models.pop(i))
 
     if len(unloaded_model) > 0:
         soft_empty_cache()
@@ -378,16 +448,17 @@ def free_memory(memory_required, device, keep_loaded=[]):
             mem_free_total, mem_free_torch = get_free_memory(device, torch_free_too=True)
             if mem_free_torch > mem_free_total * 0.25:
                 soft_empty_cache()
+    return unloaded_models
 
-def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None):
+def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
     global vram_state
 
     inference_memory = minimum_inference_memory()
-    extra_mem = max(inference_memory, memory_required)
+    extra_mem = max(inference_memory, memory_required + extra_reserved_memory())
     if minimum_memory_required is None:
         minimum_memory_required = extra_mem
     else:
-        minimum_memory_required = max(inference_memory, minimum_memory_required)
+        minimum_memory_required = max(inference_memory, minimum_memory_required + extra_reserved_memory())
 
     models = set(models)
 
@@ -420,25 +491,36 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
         devs = set(map(lambda a: a.device, models_already_loaded))
         for d in devs:
             if d != torch.device("cpu"):
-                free_memory(extra_mem, d, models_already_loaded)
+                free_memory(extra_mem + offloaded_memory(models_already_loaded, d), d, models_already_loaded)
-        return
+                free_mem = get_free_memory(d)
+                if free_mem < minimum_memory_required:
+                    logging.info("Unloading models for lowram load.") #TODO: partial model unloading when this case happens, also handle the opposite case where models can be unlowvramed.
+                    models_to_load = free_memory(minimum_memory_required, d)
+                    logging.info("{} models unloaded.".format(len(models_to_load)))
+                else:
+                    use_more_memory(free_mem - minimum_memory_required, models_already_loaded, d)
+        if len(models_to_load) == 0:
+            return
 
     logging.info(f"Loading {len(models_to_load)} new model{'s' if len(models_to_load) > 1 else ''}")
 
     total_memory_required = {}
     for loaded_model in models_to_load:
-        if unload_model_clones(loaded_model.model, unload_weights_only=True, force_unload=False) == True:#unload clones where the weights are different
+        unload_model_clones(loaded_model.model, unload_weights_only=True, force_unload=False) #unload clones where the weights are different
-            total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
+        total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
 
-    for device in total_memory_required:
+    for loaded_model in models_already_loaded:
-        if device != torch.device("cpu"):
+        total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
-            free_memory(total_memory_required[device] * 1.3 + extra_mem, device, models_already_loaded)
 
     for loaded_model in models_to_load:
         weights_unloaded = unload_model_clones(loaded_model.model, unload_weights_only=False, force_unload=False) #unload the rest of the clones where the weights can stay loaded
         if weights_unloaded is not None:
             loaded_model.weights_loaded = not weights_unloaded
 
+    for device in total_memory_required:
+        if device != torch.device("cpu"):
+            free_memory(total_memory_required[device] * 1.1 + extra_mem, device, models_already_loaded)
+
     for loaded_model in models_to_load:
         model = loaded_model.model
         torch_dev = model.load_device
@@ -447,10 +529,10 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
         else:
             vram_set_state = vram_state
         lowvram_model_memory = 0
-        if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM):
+        if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM) and not force_full_load:
             model_size = loaded_model.model_memory_required(torch_dev)
             current_free_mem = get_free_memory(torch_dev)
-            lowvram_model_memory = int(max(64 * (1024 * 1024), (current_free_mem - minimum_memory_required)))
+            lowvram_model_memory = max(64 * (1024 * 1024), (current_free_mem - minimum_memory_required), min(current_free_mem * 0.4, current_free_mem - minimum_inference_memory()))
             if model_size <= lowvram_model_memory: #only switch to lowvram if really necessary
                 lowvram_model_memory = 0
 
@@ -459,6 +541,14 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
 
         cur_loaded_model = loaded_model.model_load(lowvram_model_memory, force_patch_weights=force_patch_weights)
         current_loaded_models.insert(0, loaded_model)
+
+
+    devs = set(map(lambda a: a.device, models_already_loaded))
+    for d in devs:
+        if d != torch.device("cpu"):
+            free_mem = get_free_memory(d)
+            if free_mem > minimum_memory_required:
+                use_more_memory(free_mem - minimum_memory_required, models_already_loaded, d)
     return
 
 
@@ -478,7 +568,9 @@ def loaded_models(only_currently_used=False):
 def cleanup_models(keep_clone_weights_loaded=False):
     to_delete = []
     for i in range(len(current_loaded_models)):
-        if sys.getrefcount(current_loaded_models[i].model) <= 2:
+        #TODO: very fragile function needs improvement
+        num_refs = sys.getrefcount(current_loaded_models[i].model)
+        if num_refs <= 2:
             if not keep_clone_weights_loaded:
                 to_delete = [i] + to_delete
             #TODO: find a less fragile way to do this.
@@ -527,6 +619,9 @@ def unet_inital_load_device(parameters, dtype):
     else:
         return cpu_dev
 
+def maximum_vram_for_weights(device=None):
+    return (get_total_memory(device) * 0.88 - minimum_inference_memory())
+
 def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, torch.bfloat16, torch.float32]):
     if args.bf16_unet:
         return torch.bfloat16
@@ -536,12 +631,37 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
         return torch.float8_e4m3fn
     if args.fp8_e5m2_unet:
         return torch.float8_e5m2
-    if should_use_fp16(device=device, model_params=model_params, manual_cast=True):
+
-        if torch.float16 in supported_dtypes:
+    fp8_dtype = None
-            return torch.float16
+    try:
-    if should_use_bf16(device, model_params=model_params, manual_cast=True):
+        for dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
-        if torch.bfloat16 in supported_dtypes:
+            if dtype in supported_dtypes:
-            return torch.bfloat16
+                fp8_dtype = dtype
+                break
+    except:
+        pass
+
+    if fp8_dtype is not None:
+        free_model_memory = maximum_vram_for_weights(device)
+        if model_params * 2 > free_model_memory:
+            return fp8_dtype
+
+    for dt in supported_dtypes:
+        if dt == torch.float16 and should_use_fp16(device=device, model_params=model_params):
+            if torch.float16 in supported_dtypes:
+                return torch.float16
+        if dt == torch.bfloat16 and should_use_bf16(device, model_params=model_params):
+            if torch.bfloat16 in supported_dtypes:
+                return torch.bfloat16
+
+    for dt in supported_dtypes:
+        if dt == torch.float16 and should_use_fp16(device=device, model_params=model_params, manual_cast=True):
+            if torch.float16 in supported_dtypes:
+                return torch.float16
+        if dt == torch.bfloat16 and should_use_bf16(device, model_params=model_params, manual_cast=True):
+            if torch.bfloat16 in supported_dtypes:
+                return torch.bfloat16
+
     return torch.float32
 
 # None means no manual cast
@@ -557,13 +677,14 @@ def unet_manual_cast(weight_dtype, inference_device, supported_dtypes=[torch.flo
     if bf16_supported and weight_dtype == torch.bfloat16:
         return None
 
-    if fp16_supported and torch.float16 in supported_dtypes:
+    fp16_supported = should_use_fp16(inference_device, prioritize_performance=True)
-        return torch.float16
+    for dt in supported_dtypes:
+        if dt == torch.float16 and fp16_supported:
+            return torch.float16
+        if dt == torch.bfloat16 and bf16_supported:
+            return torch.bfloat16
 
-    elif bf16_supported and torch.bfloat16 in supported_dtypes:
+    return torch.float32
-        return torch.bfloat16
-    else:
-        return torch.float32
 
 def text_encoder_offload_device():
     if args.gpu_only:
@@ -582,6 +703,20 @@ def text_encoder_device():
     else:
         return torch.device("cpu")
 
+def text_encoder_initial_device(load_device, offload_device, model_size=0):
+    if load_device == offload_device or model_size <= 1024 * 1024 * 1024:
+        return offload_device
+
+    if is_device_mps(load_device):
+        return offload_device
+
+    mem_l = get_free_memory(load_device)
+    mem_o = get_free_memory(offload_device)
+    if mem_l > (mem_o * 0.5) and model_size * 1.2 < mem_l:
+        return load_device
+    else:
+        return offload_device
+
 def text_encoder_dtype(device=None):
     if args.fp8_e4m3fn_text_enc:
         return torch.float8_e4m3fn
@@ -758,7 +893,8 @@ def pytorch_attention_flash_attention():
 def force_upcast_attention_dtype():
     upcast = args.force_upcast_attention
     try:
-        if platform.mac_ver()[0] in ['14.5']: #black image bug on OSX Sonoma 14.5
+        macos_version = tuple(int(n) for n in platform.mac_ver()[0].split("."))
+        if (14, 5) <= macos_version < (14, 7):  # black image bug on recent versions of MacOS
             upcast = True
     except:
         pass
@@ -854,24 +990,21 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
     if torch.version.hip:
         return True
 
-    props = torch.cuda.get_device_properties("cuda")
+    props = torch.cuda.get_device_properties(device)
     if props.major >= 8:
         return True
 
     if props.major < 6:
         return False
 
-    fp16_works = False
+    #FP16 is confirmed working on a 1080 (GP104) and on latest pytorch actually seems faster than fp32
-    #FP16 is confirmed working on a 1080 (GP104) but it's a bit slower than FP32 so it should only be enabled
-    #when the model doesn't actually fit on the card
-    #TODO: actually test if GP106 and others have the same type of behavior
     nvidia_10_series = ["1080", "1070", "titan x", "p3000", "p3200", "p4000", "p4200", "p5000", "p5200", "p6000", "1060", "1050", "p40", "p100", "p6", "p4"]
     for x in nvidia_10_series:
         if x in props.name.lower():
-            fp16_works = True
+            return True
 
-    if fp16_works or manual_cast:
+    if manual_cast:
-        free_model_memory = (get_free_memory() * 0.9 - minimum_inference_memory())
+        free_model_memory = maximum_vram_for_weights(device)
         if (not prioritize_performance) or model_params * 4 > free_model_memory:
             return True
 
@@ -910,9 +1043,6 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
     if is_intel_xpu():
         return True
 
-    if device is None:
-        device = torch.device("cuda")
-
     props = torch.cuda.get_device_properties(device)
     if props.major >= 8:
         return True
@@ -920,12 +1050,22 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
     bf16_works = torch.cuda.is_bf16_supported()
 
     if bf16_works or manual_cast:
-        free_model_memory = (get_free_memory() * 0.9 - minimum_inference_memory())
+        free_model_memory = maximum_vram_for_weights(device)
         if (not prioritize_performance) or model_params * 4 > free_model_memory:
             return True
 
     return False
 
+def supports_fp8_compute(device=None):
+    props = torch.cuda.get_device_properties(device)
+    if props.major >= 9:
+        return True
+    if props.major < 8:
+        return False
+    if props.minor < 9:
+        return False
+    return True
+
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:

comfy/model_patcher.py

@@ -1,34 +1,47 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
 import torch
 import copy
 import inspect
 import logging
 import uuid
+import collections
+import math
 
 import comfy.utils
+import comfy.float
 import comfy.model_management
+import comfy.lora
 from comfy.types import UnetWrapperFunction
 
-
+def string_to_seed(data):
-def weight_decompose(dora_scale, weight, lora_diff, alpha, strength):
+    crc = 0xFFFFFFFF
-    dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, torch.float32)
+    for byte in data:
-    lora_diff *= alpha
+        if isinstance(byte, str):
-    weight_calc = weight + lora_diff.type(weight.dtype)
+            byte = ord(byte)
-    weight_norm = (
+        crc ^= byte
-        weight_calc.transpose(0, 1)
+        for _ in range(8):
-        .reshape(weight_calc.shape[1], -1)
+            if crc & 1:
-        .norm(dim=1, keepdim=True)
+                crc = (crc >> 1) ^ 0xEDB88320
-        .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1))
+            else:
-        .transpose(0, 1)
+                crc >>= 1
-    )
+    return crc ^ 0xFFFFFFFF
-
-    weight_calc *= (dora_scale / weight_norm).type(weight.dtype)
-    if strength != 1.0:
-        weight_calc -= weight
-        weight += strength * (weight_calc)
-    else:
-        weight[:] = weight_calc
-    return weight
-
 
 def set_model_options_patch_replace(model_options, patch, name, block_name, number, transformer_index=None):
     to = model_options["transformer_options"].copy()
@@ -63,10 +76,30 @@ def set_model_options_pre_cfg_function(model_options, pre_cfg_function, disable_
         model_options["disable_cfg1_optimization"] = True
     return model_options
 
+def wipe_lowvram_weight(m):
+    if hasattr(m, "prev_comfy_cast_weights"):
+        m.comfy_cast_weights = m.prev_comfy_cast_weights
+        del m.prev_comfy_cast_weights
+    m.weight_function = None
+    m.bias_function = None
+
+class LowVramPatch:
+    def __init__(self, key, patches):
+        self.key = key
+        self.patches = patches
+    def __call__(self, weight):
+        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
+
 class ModelPatcher:
-    def __init__(self, model, load_device, offload_device, size=0, current_device=None, weight_inplace_update=False):
+    def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False):
         self.size = size
         self.model = model
+        if not hasattr(self.model, 'device'):
+            logging.debug("Model doesn't have a device attribute.")
+            self.model.device = offload_device
+        elif self.model.device is None:
+            self.model.device = offload_device
+
         self.patches = {}
         self.backup = {}
         self.object_patches = {}
@@ -75,24 +108,32 @@ class ModelPatcher:
         self.model_size()
         self.load_device = load_device
         self.offload_device = offload_device
-        if current_device is None:
-            self.current_device = self.offload_device
-        else:
-            self.current_device = current_device
-
         self.weight_inplace_update = weight_inplace_update
-        self.model_lowvram = False
-        self.lowvram_patch_counter = 0
         self.patches_uuid = uuid.uuid4()
 
+        if not hasattr(self.model, 'model_loaded_weight_memory'):
+            self.model.model_loaded_weight_memory = 0
+
+        if not hasattr(self.model, 'lowvram_patch_counter'):
+            self.model.lowvram_patch_counter = 0
+
+        if not hasattr(self.model, 'model_lowvram'):
+            self.model.model_lowvram = False
+
     def model_size(self):
         if self.size > 0:
             return self.size
         self.size = comfy.model_management.module_size(self.model)
         return self.size
 
+    def loaded_size(self):
+        return self.model.model_loaded_weight_memory
+
+    def lowvram_patch_counter(self):
+        return self.model.lowvram_patch_counter
+
     def clone(self):
-        n = ModelPatcher(self.model, self.load_device, self.offload_device, self.size, self.current_device, weight_inplace_update=self.weight_inplace_update)
+        n = ModelPatcher(self.model, self.load_device, self.offload_device, self.size, weight_inplace_update=self.weight_inplace_update)
         n.patches = {}
         for k in self.patches:
             n.patches[k] = self.patches[k][:]
@@ -264,67 +305,52 @@ class ModelPatcher:
                     sd.pop(k)
         return sd
 
-    def patch_weight_to_device(self, key, device_to=None):
+    def patch_weight_to_device(self, key, device_to=None, inplace_update=False):
         if key not in self.patches:
             return
 
         weight = comfy.utils.get_attr(self.model, key)
 
-        inplace_update = self.weight_inplace_update
+        inplace_update = self.weight_inplace_update or inplace_update
 
         if key not in self.backup:
-            self.backup[key] = weight.to(device=self.offload_device, copy=inplace_update)
+            self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update)
 
         if device_to is not None:
             temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True)
         else:
             temp_weight = weight.to(torch.float32, copy=True)
-        out_weight = self.calculate_weight(self.patches[key], temp_weight, key).to(weight.dtype)
+        out_weight = comfy.lora.calculate_weight(self.patches[key], temp_weight, key)
+        out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=string_to_seed(key))
         if inplace_update:
             comfy.utils.copy_to_param(self.model, key, out_weight)
         else:
             comfy.utils.set_attr_param(self.model, key, out_weight)
 
-    def patch_model(self, device_to=None, patch_weights=True):
+    def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False):
-        for k in self.object_patches:
-            old = comfy.utils.set_attr(self.model, k, self.object_patches[k])
-            if k not in self.object_patches_backup:
-                self.object_patches_backup[k] = old
-
-        if patch_weights:
-            model_sd = self.model_state_dict()
-            for key in self.patches:
-                if key not in model_sd:
-                    logging.warning("could not patch. key doesn't exist in model: {}".format(key))
-                    continue
-
-                self.patch_weight_to_device(key, device_to)
-
-            if device_to is not None:
-                self.model.to(device_to)
-                self.current_device = device_to
-
-        return self.model
-
-    def patch_model_lowvram(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False):
-        self.patch_model(device_to, patch_weights=False)
-
-        logging.info("loading in lowvram mode {}".format(lowvram_model_memory/(1024 * 1024)))
-        class LowVramPatch:
-            def __init__(self, key, model_patcher):
-                self.key = key
-                self.model_patcher = model_patcher
-            def __call__(self, weight):
-                return self.model_patcher.calculate_weight(self.model_patcher.patches[self.key], weight, self.key)
-
         mem_counter = 0
         patch_counter = 0
+        lowvram_counter = 0
+        loading = []
         for n, m in self.model.named_modules():
+            if hasattr(m, "comfy_cast_weights") or hasattr(m, "weight"):
+                loading.append((comfy.model_management.module_size(m), n, m))
+
+        load_completely = []
+        loading.sort(reverse=True)
+        for x in loading:
+            n = x[1]
+            m = x[2]
+            module_mem = x[0]
+
             lowvram_weight = False
-            if hasattr(m, "comfy_cast_weights"):
+
-                module_mem = comfy.model_management.module_size(m)
+            if not full_load and hasattr(m, "comfy_cast_weights"):
                 if mem_counter + module_mem >= lowvram_model_memory:
                     lowvram_weight = True
+                    lowvram_counter += 1
+                    if hasattr(m, "prev_comfy_cast_weights"): #Already lowvramed
+                        continue
 
             weight_key = "{}.weight".format(n)
             bias_key = "{}.bias".format(n)
@@ -334,227 +360,173 @@ class ModelPatcher:
                     if force_patch_weights:
                         self.patch_weight_to_device(weight_key)
                     else:
-                        m.weight_function = LowVramPatch(weight_key, self)
+                        m.weight_function = LowVramPatch(weight_key, self.patches)
                         patch_counter += 1
                 if bias_key in self.patches:
                     if force_patch_weights:
                         self.patch_weight_to_device(bias_key)
                     else:
-                        m.bias_function = LowVramPatch(bias_key, self)
+                        m.bias_function = LowVramPatch(bias_key, self.patches)
                         patch_counter += 1
 
                 m.prev_comfy_cast_weights = m.comfy_cast_weights
                 m.comfy_cast_weights = True
             else:
+                if hasattr(m, "comfy_cast_weights"):
+                    if m.comfy_cast_weights:
+                        wipe_lowvram_weight(m)
+
                 if hasattr(m, "weight"):
-                    self.patch_weight_to_device(weight_key, device_to)
+                    mem_counter += module_mem
-                    self.patch_weight_to_device(bias_key, device_to)
+                    load_completely.append((module_mem, n, m))
-                    m.to(device_to)
-                    mem_counter += comfy.model_management.module_size(m)
-                    logging.debug("lowvram: loaded module regularly {} {}".format(n, m))
 
-        self.model_lowvram = True
+        load_completely.sort(reverse=True)
-        self.lowvram_patch_counter = patch_counter
+        for x in load_completely:
+            n = x[1]
+            m = x[2]
+            weight_key = "{}.weight".format(n)
+            bias_key = "{}.bias".format(n)
+            if hasattr(m, "comfy_patched_weights"):
+                if m.comfy_patched_weights == True:
+                    continue
+
+            self.patch_weight_to_device(weight_key, device_to=device_to)
+            self.patch_weight_to_device(bias_key, device_to=device_to)
+            logging.debug("lowvram: loaded module regularly {} {}".format(n, m))
+            m.comfy_patched_weights = True
+
+        for x in load_completely:
+            x[2].to(device_to)
+
+        if lowvram_counter > 0:
+            logging.info("loaded partially {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), patch_counter))
+            self.model.model_lowvram = True
+        else:
+            logging.info("loaded completely {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
+            self.model.model_lowvram = False
+            if full_load:
+                self.model.to(device_to)
+                mem_counter = self.model_size()
+
+        self.model.lowvram_patch_counter += patch_counter
+        self.model.device = device_to
+        self.model.model_loaded_weight_memory = mem_counter
+
+    def patch_model(self, device_to=None, lowvram_model_memory=0, load_weights=True, force_patch_weights=False):
+        for k in self.object_patches:
+            old = comfy.utils.set_attr(self.model, k, self.object_patches[k])
+            if k not in self.object_patches_backup:
+                self.object_patches_backup[k] = old
+
+        if lowvram_model_memory == 0:
+            full_load = True
+        else:
+            full_load = False
+
+        if load_weights:
+            self.load(device_to, lowvram_model_memory=lowvram_model_memory, force_patch_weights=force_patch_weights, full_load=full_load)
         return self.model
 
-    def calculate_weight(self, patches, weight, key):
-        for p in patches:
-            strength = p[0]
-            v = p[1]
-            strength_model = p[2]
-            offset = p[3]
-            function = p[4]
-            if function is None:
-                function = lambda a: a
-
-            old_weight = None
-            if offset is not None:
-                old_weight = weight
-                weight = weight.narrow(offset[0], offset[1], offset[2])
-
-            if strength_model != 1.0:
-                weight *= strength_model
-
-            if isinstance(v, list):
-                v = (self.calculate_weight(v[1:], v[0].clone(), key), )
-
-            if len(v) == 1:
-                patch_type = "diff"
-            elif len(v) == 2:
-                patch_type = v[0]
-                v = v[1]
-
-            if patch_type == "diff":
-                w1 = v[0]
-                if strength != 0.0:
-                    if w1.shape != weight.shape:
-                        logging.warning("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED {} != {}".format(key, w1.shape, weight.shape))
-                    else:
-                        weight += function(strength * comfy.model_management.cast_to_device(w1, weight.device, weight.dtype))
-            elif patch_type == "lora": #lora/locon
-                mat1 = comfy.model_management.cast_to_device(v[0], weight.device, torch.float32)
-                mat2 = comfy.model_management.cast_to_device(v[1], weight.device, torch.float32)
-                dora_scale = v[4]
-                if v[2] is not None:
-                    alpha = v[2] / mat2.shape[0]
-                else:
-                    alpha = 1.0
-
-                if v[3] is not None:
-                    #locon mid weights, hopefully the math is fine because I didn't properly test it
-                    mat3 = comfy.model_management.cast_to_device(v[3], weight.device, torch.float32)
-                    final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
-                    mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1)
-                try:
-                    lora_diff = torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)).reshape(weight.shape)
-                    if dora_scale is not None:
-                        weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength))
-                    else:
-                        weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-                except Exception as e:
-                    logging.error("ERROR {} {} {}".format(patch_type, key, e))
-            elif patch_type == "lokr":
-                w1 = v[0]
-                w2 = v[1]
-                w1_a = v[3]
-                w1_b = v[4]
-                w2_a = v[5]
-                w2_b = v[6]
-                t2 = v[7]
-                dora_scale = v[8]
-                dim = None
-
-                if w1 is None:
-                    dim = w1_b.shape[0]
-                    w1 = torch.mm(comfy.model_management.cast_to_device(w1_a, weight.device, torch.float32),
-                                  comfy.model_management.cast_to_device(w1_b, weight.device, torch.float32))
-                else:
-                    w1 = comfy.model_management.cast_to_device(w1, weight.device, torch.float32)
-
-                if w2 is None:
-                    dim = w2_b.shape[0]
-                    if t2 is None:
-                        w2 = torch.mm(comfy.model_management.cast_to_device(w2_a, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w2_b, weight.device, torch.float32))
-                    else:
-                        w2 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                          comfy.model_management.cast_to_device(t2, weight.device, torch.float32),
-                                          comfy.model_management.cast_to_device(w2_b, weight.device, torch.float32),
-                                          comfy.model_management.cast_to_device(w2_a, weight.device, torch.float32))
-                else:
-                    w2 = comfy.model_management.cast_to_device(w2, weight.device, torch.float32)
-
-                if len(w2.shape) == 4:
-                    w1 = w1.unsqueeze(2).unsqueeze(2)
-                if v[2] is not None and dim is not None:
-                    alpha = v[2] / dim
-                else:
-                    alpha = 1.0
-
-                try:
-                    lora_diff = torch.kron(w1, w2).reshape(weight.shape)
-                    if dora_scale is not None:
-                        weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength))
-                    else:
-                        weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-                except Exception as e:
-                    logging.error("ERROR {} {} {}".format(patch_type, key, e))
-            elif patch_type == "loha":
-                w1a = v[0]
-                w1b = v[1]
-                if v[2] is not None:
-                    alpha = v[2] / w1b.shape[0]
-                else:
-                    alpha = 1.0
-
-                w2a = v[3]
-                w2b = v[4]
-                dora_scale = v[7]
-                if v[5] is not None: #cp decomposition
-                    t1 = v[5]
-                    t2 = v[6]
-                    m1 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                      comfy.model_management.cast_to_device(t1, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w1b, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w1a, weight.device, torch.float32))
-
-                    m2 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                      comfy.model_management.cast_to_device(t2, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w2b, weight.device, torch.float32),
-                                      comfy.model_management.cast_to_device(w2a, weight.device, torch.float32))
-                else:
-                    m1 = torch.mm(comfy.model_management.cast_to_device(w1a, weight.device, torch.float32),
-                                  comfy.model_management.cast_to_device(w1b, weight.device, torch.float32))
-                    m2 = torch.mm(comfy.model_management.cast_to_device(w2a, weight.device, torch.float32),
-                                  comfy.model_management.cast_to_device(w2b, weight.device, torch.float32))
-
-                try:
-                    lora_diff = (m1 * m2).reshape(weight.shape)
-                    if dora_scale is not None:
-                        weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength))
-                    else:
-                        weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-                except Exception as e:
-                    logging.error("ERROR {} {} {}".format(patch_type, key, e))
-            elif patch_type == "glora":
-                if v[4] is not None:
-                    alpha = v[4] / v[0].shape[0]
-                else:
-                    alpha = 1.0
-
-                dora_scale = v[5]
-
-                a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, torch.float32)
-                a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, torch.float32)
-                b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, torch.float32)
-                b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, torch.float32)
-
-                try:
-                    lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1), a2), a1)).reshape(weight.shape)
-                    if dora_scale is not None:
-                        weight = function(weight_decompose(dora_scale, weight, lora_diff, alpha, strength))
-                    else:
-                        weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-                except Exception as e:
-                    logging.error("ERROR {} {} {}".format(patch_type, key, e))
-            else:
-                logging.warning("patch type not recognized {} {}".format(patch_type, key))
-
-            if old_weight is not None:
-                weight = old_weight
-
-        return weight
-
     def unpatch_model(self, device_to=None, unpatch_weights=True):
         if unpatch_weights:
-            if self.model_lowvram:
+            if self.model.model_lowvram:
                 for m in self.model.modules():
-                    if hasattr(m, "prev_comfy_cast_weights"):
+                    wipe_lowvram_weight(m)
-                        m.comfy_cast_weights = m.prev_comfy_cast_weights
-                        del m.prev_comfy_cast_weights
-                    m.weight_function = None
-                    m.bias_function = None
 
-                self.model_lowvram = False
+                self.model.model_lowvram = False
-                self.lowvram_patch_counter = 0
+                self.model.lowvram_patch_counter = 0
 
             keys = list(self.backup.keys())
 
-            if self.weight_inplace_update:
+            for k in keys:
-                for k in keys:
+                bk = self.backup[k]
-                    comfy.utils.copy_to_param(self.model, k, self.backup[k])
+                if bk.inplace_update:
-            else:
+                    comfy.utils.copy_to_param(self.model, k, bk.weight)
-                for k in keys:
+                else:
-                    comfy.utils.set_attr_param(self.model, k, self.backup[k])
+                    comfy.utils.set_attr_param(self.model, k, bk.weight)
 
             self.backup.clear()
 
             if device_to is not None:
                 self.model.to(device_to)
-                self.current_device = device_to
+                self.model.device = device_to
+            self.model.model_loaded_weight_memory = 0
+
+            for m in self.model.modules():
+                if hasattr(m, "comfy_patched_weights"):
+                    del m.comfy_patched_weights
 
         keys = list(self.object_patches_backup.keys())
         for k in keys:
             comfy.utils.set_attr(self.model, k, self.object_patches_backup[k])
 
         self.object_patches_backup.clear()
+
+    def partially_unload(self, device_to, memory_to_free=0):
+        memory_freed = 0
+        patch_counter = 0
+        unload_list = []
+
+        for n, m in self.model.named_modules():
+            shift_lowvram = False
+            if hasattr(m, "comfy_cast_weights"):
+                module_mem = comfy.model_management.module_size(m)
+                unload_list.append((module_mem, n, m))
+
+        unload_list.sort()
+        for unload in unload_list:
+            if memory_to_free < memory_freed:
+                break
+            module_mem = unload[0]
+            n = unload[1]
+            m = unload[2]
+            weight_key = "{}.weight".format(n)
+            bias_key = "{}.bias".format(n)
+
+            if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:
+                for key in [weight_key, bias_key]:
+                    bk = self.backup.get(key, None)
+                    if bk is not None:
+                        if bk.inplace_update:
+                            comfy.utils.copy_to_param(self.model, key, bk.weight)
+                        else:
+                            comfy.utils.set_attr_param(self.model, key, bk.weight)
+                        self.backup.pop(key)
+
+                m.to(device_to)
+                if weight_key in self.patches:
+                    m.weight_function = LowVramPatch(weight_key, self.patches)
+                    patch_counter += 1
+                if bias_key in self.patches:
+                    m.bias_function = LowVramPatch(bias_key, self.patches)
+                    patch_counter += 1
+
+                m.prev_comfy_cast_weights = m.comfy_cast_weights
+                m.comfy_cast_weights = True
+                m.comfy_patched_weights = False
+                memory_freed += module_mem
+                logging.debug("freed {}".format(n))
+
+        self.model.model_lowvram = True
+        self.model.lowvram_patch_counter += patch_counter
+        self.model.model_loaded_weight_memory -= memory_freed
+        return memory_freed
+
+    def partially_load(self, device_to, extra_memory=0):
+        self.unpatch_model(unpatch_weights=False)
+        self.patch_model(load_weights=False)
+        full_load = False
+        if self.model.model_lowvram == False:
+            return 0
+        if self.model.model_loaded_weight_memory + extra_memory > self.model_size():
+            full_load = True
+        current_used = self.model.model_loaded_weight_memory
+        self.load(device_to, lowvram_model_memory=current_used + extra_memory, full_load=full_load)
+        return self.model.model_loaded_weight_memory - current_used
+
+    def current_loaded_device(self):
+        return self.model.device
+
+    def calculate_weight(self, patches, weight, key, intermediate_dtype=torch.float32):
+        print("WARNING the ModelPatcher.calculate_weight function is deprecated, please use: comfy.lora.calculate_weight instead")
+        return comfy.lora.calculate_weight(patches, weight, key, intermediate_dtype=intermediate_dtype)

comfy/ops.py

@@ -18,29 +18,42 @@
 
 import torch
 import comfy.model_management
+from comfy.cli_args import args
 
+def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False):
+    if device is None or weight.device == device:
+        if not copy:
+            if dtype is None or weight.dtype == dtype:
+                return weight
+        return weight.to(dtype=dtype, copy=copy)
 
-def cast_to(weight, dtype=None, device=None, non_blocking=False):
+    r = torch.empty_like(weight, dtype=dtype, device=device)
-    return weight.to(device=device, dtype=dtype, non_blocking=non_blocking)
+    r.copy_(weight, non_blocking=non_blocking)
+    return r
 
-def cast_to_input(weight, input, non_blocking=False):
+def cast_to_input(weight, input, non_blocking=False, copy=True):
-    return cast_to(weight, input.dtype, input.device, non_blocking=non_blocking)
+    return cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 
-def cast_bias_weight(s, input=None, dtype=None, device=None):
+def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
     if input is not None:
         if dtype is None:
             dtype = input.dtype
+        if bias_dtype is None:
+            bias_dtype = dtype
         if device is None:
             device = input.device
 
     bias = None
-    non_blocking = comfy.model_management.device_should_use_non_blocking(device)
+    non_blocking = comfy.model_management.device_supports_non_blocking(device)
     if s.bias is not None:
-        bias = cast_to(s.bias, dtype, device, non_blocking=non_blocking)
+        has_function = s.bias_function is not None
-        if s.bias_function is not None:
+        bias = cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function)
+        if has_function:
             bias = s.bias_function(bias)
-    weight = cast_to(s.weight, dtype, device, non_blocking=non_blocking)
+
-    if s.weight_function is not None:
+    has_function = s.weight_function is not None
+    weight = cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function)
+    if has_function:
         weight = s.weight_function(weight)
     return weight, bias
 
@@ -238,3 +251,59 @@ class manual_cast(disable_weight_init):
 
     class Embedding(disable_weight_init.Embedding):
         comfy_cast_weights = True
+
+
+def fp8_linear(self, input):
+    dtype = self.weight.dtype
+    if dtype not in [torch.float8_e4m3fn]:
+        return None
+
+    if len(input.shape) == 3:
+        inn = input.reshape(-1, input.shape[2]).to(dtype)
+        non_blocking = comfy.model_management.device_supports_non_blocking(input.device)
+        w, bias = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input.dtype)
+        w = w.t()
+
+        scale_weight = self.scale_weight
+        scale_input = self.scale_input
+        if scale_weight is None:
+            scale_weight = torch.ones((1), device=input.device, dtype=torch.float32)
+            if scale_input is None:
+                scale_input = scale_weight
+        if scale_input is None:
+            scale_input = torch.ones((1), device=input.device, dtype=torch.float32)
+
+        if bias is not None:
+            o = torch._scaled_mm(inn, w, out_dtype=input.dtype, bias=bias, scale_a=scale_input, scale_b=scale_weight)
+        else:
+            o = torch._scaled_mm(inn, w, out_dtype=input.dtype, scale_a=scale_input, scale_b=scale_weight)
+
+        if isinstance(o, tuple):
+            o = o[0]
+
+        return o.reshape((-1, input.shape[1], self.weight.shape[0]))
+    return None
+
+class fp8_ops(manual_cast):
+    class Linear(manual_cast.Linear):
+        def reset_parameters(self):
+            self.scale_weight = None
+            self.scale_input = None
+            return None
+
+        def forward_comfy_cast_weights(self, input):
+            out = fp8_linear(self, input)
+            if out is not None:
+                return out
+
+            weight, bias = cast_bias_weight(self, input)
+            return torch.nn.functional.linear(input, weight, bias)
+
+
+def pick_operations(weight_dtype, compute_dtype, load_device=None):
+    if compute_dtype is None or weight_dtype == compute_dtype:
+        return disable_weight_init
+    if args.fast:
+        if comfy.model_management.supports_fp8_compute(load_device):
+            return fp8_ops
+    return manual_cast

comfy/samplers.py

@@ -171,7 +171,7 @@ def calc_cond_batch(model, conds, x_in, timestep, model_options):
         for i in range(1, len(to_batch_temp) + 1):
             batch_amount = to_batch_temp[:len(to_batch_temp)//i]
             input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:]
-            if model.memory_required(input_shape) < free_memory:
+            if model.memory_required(input_shape) * 1.5 < free_memory:
                 to_batch = batch_amount
                 break
 

comfy/sd.py

@@ -24,6 +24,7 @@ import comfy.text_encoders.sa_t5
 import comfy.text_encoders.aura_t5
 import comfy.text_encoders.hydit
 import comfy.text_encoders.flux
+import comfy.text_encoders.long_clipl
 
 import comfy.model_patcher
 import comfy.lora
@@ -62,7 +63,7 @@ def load_lora_for_models(model, clip, lora, strength_model, strength_clip):
 
 
 class CLIP:
-    def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}):
+    def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, model_options={}):
         if no_init:
             return
         params = target.params.copy()
@@ -71,20 +72,29 @@ class CLIP:
 
         load_device = model_management.text_encoder_device()
         offload_device = model_management.text_encoder_offload_device()
-        params['device'] = offload_device
+        dtype = model_options.get("dtype", None)
-        dtype = model_management.text_encoder_dtype(load_device)
+        if dtype is None:
+            dtype = model_management.text_encoder_dtype(load_device)
+
         params['dtype'] = dtype
+        params['device'] = model_management.text_encoder_initial_device(load_device, offload_device, parameters * model_management.dtype_size(dtype))
+        params['model_options'] = model_options
 
         self.cond_stage_model = clip(**(params))
 
         for dt in self.cond_stage_model.dtypes:
             if not model_management.supports_cast(load_device, dt):
                 load_device = offload_device
+                if params['device'] != offload_device:
+                    self.cond_stage_model.to(offload_device)
+                    logging.warning("Had to shift TE back.")
 
         self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
         self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
+        if params['device'] == load_device:
+            model_management.load_models_gpu([self.patcher], force_full_load=True)
         self.layer_idx = None
-        logging.debug("CLIP model load device: {}, offload device: {}".format(load_device, offload_device))
+        logging.debug("CLIP model load device: {}, offload device: {}, current: {}".format(load_device, offload_device, params['device']))
 
     def clone(self):
         n = CLIP(no_init=True)
@@ -390,11 +400,14 @@ class CLIPType(Enum):
     HUNYUAN_DIT = 5
     FLUX = 6
 
-def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION):
+def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
     clip_data = []
     for p in ckpt_paths:
         clip_data.append(comfy.utils.load_torch_file(p, safe_load=True))
+    return load_text_encoder_state_dicts(clip_data, embedding_directory=embedding_directory, clip_type=clip_type, model_options=model_options)
 
+def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
+    clip_data = state_dicts
     class EmptyClass:
         pass
 
@@ -431,8 +444,13 @@ def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DI
             clip_target.clip = comfy.text_encoders.sa_t5.SAT5Model
             clip_target.tokenizer = comfy.text_encoders.sa_t5.SAT5Tokenizer
         else:
-            clip_target.clip = sd1_clip.SD1ClipModel
+            w = clip_data[0].get("text_model.embeddings.position_embedding.weight", None)
-            clip_target.tokenizer = sd1_clip.SD1Tokenizer
+            if w is not None and w.shape[0] == 248:
+                clip_target.clip = comfy.text_encoders.long_clipl.LongClipModel
+                clip_target.tokenizer = comfy.text_encoders.long_clipl.LongClipTokenizer
+            else:
+                clip_target.clip = sd1_clip.SD1ClipModel
+                clip_target.tokenizer = sd1_clip.SD1Tokenizer
     elif len(clip_data) == 2:
         if clip_type == CLIPType.SD3:
             clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=True, clip_g=True, t5=False)
@@ -456,7 +474,11 @@ def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DI
         clip_target.clip = comfy.text_encoders.sd3_clip.SD3ClipModel
         clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
 
-    clip = CLIP(clip_target, embedding_directory=embedding_directory)
+    parameters = 0
+    for c in clip_data:
+        parameters += comfy.utils.calculate_parameters(c)
+
+    clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, model_options=model_options)
     for c in clip_data:
         m, u = clip.load_sd(c)
         if len(m) > 0:
@@ -498,25 +520,39 @@ def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_cl
 
     return (model, clip, vae)
 
-def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True):
+def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}):
     sd = comfy.utils.load_torch_file(ckpt_path)
-    sd_keys = sd.keys()
+    out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options)
+    if out is None:
+        raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))
+    return out
+
+def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}):
     clip = None
     clipvision = None
     vae = None
     model = None
     model_patcher = None
-    clip_target = None
 
     diffusion_model_prefix = model_detection.unet_prefix_from_state_dict(sd)
     parameters = comfy.utils.calculate_parameters(sd, diffusion_model_prefix)
+    weight_dtype = comfy.utils.weight_dtype(sd, diffusion_model_prefix)
     load_device = model_management.get_torch_device()
 
     model_config = model_detection.model_config_from_unet(sd, diffusion_model_prefix)
     if model_config is None:
-        raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))
+        return None
+
+    unet_weight_dtype = list(model_config.supported_inference_dtypes)
+    if weight_dtype is not None:
+        unet_weight_dtype.append(weight_dtype)
+
+    model_config.custom_operations = model_options.get("custom_operations", None)
+    unet_dtype = model_options.get("weight_dtype", None)
+
+    if unet_dtype is None:
+        unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=unet_weight_dtype)
 
-    unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=model_config.supported_inference_dtypes)
     manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
     model_config.set_inference_dtype(unet_dtype, manual_cast_dtype)
 
@@ -540,7 +576,8 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
         if clip_target is not None:
             clip_sd = model_config.process_clip_state_dict(sd)
             if len(clip_sd) > 0:
-                clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd)
+                parameters = comfy.utils.calculate_parameters(clip_sd)
+                clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, model_options=te_model_options)
                 m, u = clip.load_sd(clip_sd, full_model=True)
                 if len(m) > 0:
                     m_filter = list(filter(lambda a: ".logit_scale" not in a and ".transformer.text_projection.weight" not in a, m))
@@ -559,15 +596,16 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
         logging.debug("left over keys: {}".format(left_over))
 
     if output_model:
-        model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device(), current_device=inital_load_device)
+        model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device())
         if inital_load_device != torch.device("cpu"):
             logging.info("loaded straight to GPU")
-            model_management.load_model_gpu(model_patcher)
+            model_management.load_models_gpu([model_patcher], force_full_load=True)
 
     return (model_patcher, clip, vae, clipvision)
 
 
-def load_unet_state_dict(sd, dtype=None): #load unet in diffusers or regular format
+def load_diffusion_model_state_dict(sd, model_options={}): #load unet in diffusers or regular format
+    dtype = model_options.get("dtype", None)
 
     #Allow loading unets from checkpoint files
     diffusion_model_prefix = model_detection.unet_prefix_from_state_dict(sd)
@@ -609,6 +647,7 @@ def load_unet_state_dict(sd, dtype=None): #load unet in diffusers or regular for
 
     manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
     model_config.set_inference_dtype(unet_dtype, manual_cast_dtype)
+    model_config.custom_operations = model_options.get("custom_operations", None)
     model = model_config.get_model(new_sd, "")
     model = model.to(offload_device)
     model.load_model_weights(new_sd, "")
@@ -617,24 +656,36 @@ def load_unet_state_dict(sd, dtype=None): #load unet in diffusers or regular for
         logging.info("left over keys in unet: {}".format(left_over))
     return comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device)
 
-def load_unet(unet_path, dtype=None):
+
+def load_diffusion_model(unet_path, model_options={}):
     sd = comfy.utils.load_torch_file(unet_path)
-    model = load_unet_state_dict(sd, dtype=dtype)
+    model = load_diffusion_model_state_dict(sd, model_options=model_options)
     if model is None:
         logging.error("ERROR UNSUPPORTED UNET {}".format(unet_path))
         raise RuntimeError("ERROR: Could not detect model type of: {}".format(unet_path))
     return model
 
+def load_unet(unet_path, dtype=None):
+    print("WARNING: the load_unet function has been deprecated and will be removed please switch to: load_diffusion_model")
+    return load_diffusion_model(unet_path, model_options={"dtype": dtype})
+
+def load_unet_state_dict(sd, dtype=None):
+    print("WARNING: the load_unet_state_dict function has been deprecated and will be removed please switch to: load_diffusion_model_state_dict")
+    return load_diffusion_model_state_dict(sd, model_options={"dtype": dtype})
+
 def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, metadata=None, extra_keys={}):
     clip_sd = None
     load_models = [model]
     if clip is not None:
         load_models.append(clip.load_model())
         clip_sd = clip.get_sd()
+    vae_sd = None
+    if vae is not None:
+        vae_sd = vae.get_sd()
 
     model_management.load_models_gpu(load_models, force_patch_weights=True)
     clip_vision_sd = clip_vision.get_sd() if clip_vision is not None else None
-    sd = model.model.state_dict_for_saving(clip_sd, vae.get_sd(), clip_vision_sd)
+    sd = model.model.state_dict_for_saving(clip_sd, vae_sd, clip_vision_sd)
     for k in extra_keys:
         sd[k] = extra_keys[k]
 

comfy/sd1_clip.py

@@ -75,7 +75,6 @@ class ClipTokenWeightEncoder:
         return r
 
 class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
-    """Uses the CLIP transformer encoder for text (from huggingface)"""
     LAYERS = [
         "last",
         "pooled",
@@ -84,7 +83,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
     def __init__(self, version="openai/clip-vit-large-patch14", device="cpu", max_length=77,
                  freeze=True, layer="last", layer_idx=None, textmodel_json_config=None, dtype=None, model_class=comfy.clip_model.CLIPTextModel,
                  special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True, enable_attention_masks=False, zero_out_masked=False,
-                 return_projected_pooled=True, return_attention_masks=False):  # clip-vit-base-patch32
+                 return_projected_pooled=True, return_attention_masks=False, model_options={}):  # clip-vit-base-patch32
         super().__init__()
         assert layer in self.LAYERS
 
@@ -94,7 +93,11 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         with open(textmodel_json_config) as f:
             config = json.load(f)
 
-        self.operations = comfy.ops.manual_cast
+        operations = model_options.get("custom_operations", None)
+        if operations is None:
+            operations = comfy.ops.manual_cast
+
+        self.operations = operations
         self.transformer = model_class(config, dtype, device, self.operations)
         self.num_layers = self.transformer.num_layers
 
@@ -313,6 +316,17 @@ def expand_directory_list(directories):
             dirs.add(root)
     return list(dirs)
 
+def bundled_embed(embed, prefix, suffix): #bundled embedding in lora format
+    i = 0
+    out_list = []
+    for k in embed:
+        if k.startswith(prefix) and k.endswith(suffix):
+            out_list.append(embed[k])
+    if len(out_list) == 0:
+        return None
+
+    return torch.cat(out_list, dim=0)
+
 def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=None):
     if isinstance(embedding_directory, str):
         embedding_directory = [embedding_directory]
@@ -379,8 +393,12 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No
         elif embed_key is not None and embed_key in embed:
             embed_out = embed[embed_key]
         else:
-            values = embed.values()
+            embed_out = bundled_embed(embed, 'bundle_emb.', '.string_to_param.*')
-            embed_out = next(iter(values))
+            if embed_out is None:
+                embed_out = bundled_embed(embed, 'bundle_emb.', '.{}'.format(embed_key))
+            if embed_out is None:
+                values = embed.values()
+                embed_out = next(iter(values))
     return embed_out
 
 class SDTokenizer:
@@ -537,8 +555,12 @@ class SD1Tokenizer:
     def state_dict(self):
         return {}
 
+class SD1CheckpointClipModel(SDClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, return_projected_pooled=False, dtype=dtype, model_options=model_options)
+
 class SD1ClipModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None, clip_name="l", clip_model=SDClipModel, name=None, **kwargs):
+    def __init__(self, device="cpu", dtype=None, model_options={}, clip_name="l", clip_model=SD1CheckpointClipModel, name=None, **kwargs):
         super().__init__()
 
         if name is not None:
@@ -548,7 +570,7 @@ class SD1ClipModel(torch.nn.Module):
             self.clip_name = clip_name
             self.clip = "clip_{}".format(self.clip_name)
 
-        setattr(self, self.clip, clip_model(device=device, dtype=dtype, **kwargs))
+        setattr(self, self.clip, clip_model(device=device, dtype=dtype, model_options=model_options, **kwargs))
 
         self.dtypes = set()
         if dtype is not None:

comfy/sdxl_clip.py

@@ -3,14 +3,14 @@ import torch
 import os
 
 class SDXLClipG(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None, model_options={}):
         if layer == "penultimate":
             layer="hidden"
             layer_idx=-2
 
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
         super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
-                         special_tokens={"start": 49406, "end": 49407, "pad": 0}, layer_norm_hidden_state=False)
+                         special_tokens={"start": 49406, "end": 49407, "pad": 0}, layer_norm_hidden_state=False, return_projected_pooled=True, model_options=model_options)
 
     def load_sd(self, sd):
         return super().load_sd(sd)
@@ -38,10 +38,10 @@ class SDXLTokenizer:
         return {}
 
 class SDXLClipModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
         super().__init__()
-        self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False)
+        self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, model_options=model_options)
-        self.clip_g = SDXLClipG(device=device, dtype=dtype)
+        self.clip_g = SDXLClipG(device=device, dtype=dtype, model_options=model_options)
         self.dtypes = set([dtype])
 
     def set_clip_options(self, options):
@@ -66,8 +66,8 @@ class SDXLClipModel(torch.nn.Module):
             return self.clip_l.load_sd(sd)
 
 class SDXLRefinerClipModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
-        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=SDXLClipG)
+        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=SDXLClipG, model_options=model_options)
 
 
 class StableCascadeClipGTokenizer(sd1_clip.SDTokenizer):
@@ -79,14 +79,14 @@ class StableCascadeTokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="g", tokenizer=StableCascadeClipGTokenizer)
 
 class StableCascadeClipG(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", max_length=77, freeze=True, layer="hidden", layer_idx=-1, dtype=None):
+    def __init__(self, device="cpu", max_length=77, freeze=True, layer="hidden", layer_idx=-1, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
         super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
-                         special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=False, enable_attention_masks=True)
+                         special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=False, enable_attention_masks=True, return_projected_pooled=True, model_options=model_options)
 
     def load_sd(self, sd):
         return super().load_sd(sd)
 
 class StableCascadeClipModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
-        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=StableCascadeClipG)
+        super().__init__(device=device, dtype=dtype, clip_name="g", clip_model=StableCascadeClipG, model_options=model_options)

comfy/supported_models.py

@@ -31,6 +31,7 @@ class SD15(supported_models_base.BASE):
     }
 
     latent_format = latent_formats.SD15
+    memory_usage_factor = 1.0
 
     def process_clip_state_dict(self, state_dict):
         k = list(state_dict.keys())
@@ -77,6 +78,7 @@ class SD20(supported_models_base.BASE):
     }
 
     latent_format = latent_formats.SD15
+    memory_usage_factor = 1.0
 
     def model_type(self, state_dict, prefix=""):
         if self.unet_config["in_channels"] == 4: #SD2.0 inpainting models are not v prediction
@@ -140,6 +142,7 @@ class SDXLRefiner(supported_models_base.BASE):
     }
 
     latent_format = latent_formats.SDXL
+    memory_usage_factor = 1.0
 
     def get_model(self, state_dict, prefix="", device=None):
         return model_base.SDXLRefiner(self, device=device)
@@ -178,6 +181,8 @@ class SDXL(supported_models_base.BASE):
 
     latent_format = latent_formats.SDXL
 
+    memory_usage_factor = 0.8
+
     def model_type(self, state_dict, prefix=""):
         if 'edm_mean' in state_dict and 'edm_std' in state_dict: #Playground V2.5
             self.latent_format = latent_formats.SDXL_Playground_2_5()
@@ -505,6 +510,9 @@ class SD3(supported_models_base.BASE):
 
     unet_extra_config = {}
     latent_format = latent_formats.SD3
+
+    memory_usage_factor = 1.2
+
     text_encoder_key_prefix = ["text_encoders."]
 
     def get_model(self, state_dict, prefix="", device=None):
@@ -631,7 +639,10 @@ class Flux(supported_models_base.BASE):
 
     unet_extra_config = {}
     latent_format = latent_formats.Flux
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    memory_usage_factor = 2.8
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
     vae_key_prefix = ["vae."]
     text_encoder_key_prefix = ["text_encoders."]
@@ -641,7 +652,12 @@ class Flux(supported_models_base.BASE):
         return out
 
     def clip_target(self, state_dict={}):
-        return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.FluxClipModel)
+        pref = self.text_encoder_key_prefix[0]
+        t5_key = "{}t5xxl.transformer.encoder.final_layer_norm.weight".format(pref)
+        dtype_t5 = None
+        if t5_key in state_dict:
+            dtype_t5 = state_dict[t5_key].dtype
+        return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.flux_clip(dtype_t5=dtype_t5))
 
 class FluxSchnell(Flux):
     unet_config = {

comfy/supported_models_base.py

@@ -1,3 +1,21 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
 import torch
 from . import model_base
 from . import utils
@@ -27,7 +45,10 @@ class BASE:
     text_encoder_key_prefix = ["cond_stage_model."]
     supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32]
 
+    memory_usage_factor = 2.0
+
     manual_cast_dtype = None
+    custom_operations = None
 
     @classmethod
     def matches(s, unet_config, state_dict=None):

comfy/text_encoders/aura_t5.py

@@ -4,9 +4,9 @@ import comfy.text_encoders.t5
 import os
 
 class PT5XlModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_pile_config_xl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 2, "pad": 1}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 2, "pad": 1}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True, model_options=model_options)
 
 class PT5XlTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -18,5 +18,5 @@ class AuraT5Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="pile_t5xl", tokenizer=PT5XlTokenizer)
 
 class AuraT5Model(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, **kwargs):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
-        super().__init__(device=device, dtype=dtype, name="pile_t5xl", clip_model=PT5XlModel, **kwargs)
+        super().__init__(device=device, dtype=dtype, model_options=model_options, name="pile_t5xl", clip_model=PT5XlModel, **kwargs)

comfy/text_encoders/flux.py

@@ -6,9 +6,9 @@ import torch
 import os
 
 class T5XXLModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, model_options=model_options)
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -35,11 +35,11 @@ class FluxTokenizer:
 
 
 class FluxClipModel(torch.nn.Module):
-    def __init__(self, dtype_t5=None, device="cpu", dtype=None):
+    def __init__(self, dtype_t5=None, device="cpu", dtype=None, model_options={}):
         super().__init__()
         dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
-        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False)
+        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False, model_options=model_options)
-        self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5)
+        self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5, model_options=model_options)
         self.dtypes = set([dtype, dtype_t5])
 
     def set_clip_options(self, options):
@@ -52,9 +52,9 @@ class FluxClipModel(torch.nn.Module):
 
     def encode_token_weights(self, token_weight_pairs):
         token_weight_pairs_l = token_weight_pairs["l"]
-        token_weight_pars_t5 = token_weight_pairs["t5xxl"]
+        token_weight_pairs_t5 = token_weight_pairs["t5xxl"]
 
-        t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pars_t5)
+        t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pairs_t5)
         l_out, l_pooled = self.clip_l.encode_token_weights(token_weight_pairs_l)
         return t5_out, l_pooled
 
@@ -66,6 +66,6 @@ class FluxClipModel(torch.nn.Module):
 
 def flux_clip(dtype_t5=None):
     class FluxClipModel_(FluxClipModel):
-        def __init__(self, device="cpu", dtype=None):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
-            super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype)
+            super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options)
     return FluxClipModel_

comfy/text_encoders/hydit.py

@@ -7,9 +7,9 @@ import os
 import torch
 
 class HyditBertModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "hydit_clip.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 101, "end": 102, "pad": 0}, model_class=BertModel, enable_attention_masks=True, return_attention_masks=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 101, "end": 102, "pad": 0}, model_class=BertModel, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
 
 class HyditBertTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -18,9 +18,9 @@ class HyditBertTokenizer(sd1_clip.SDTokenizer):
 
 
 class MT5XLModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "mt5_config_xl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, return_attention_masks=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
 
 class MT5XLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -50,10 +50,10 @@ class HyditTokenizer:
         return {"mt5xl.spiece_model": self.mt5xl.state_dict()["spiece_model"]}
 
 class HyditModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
         super().__init__()
-        self.hydit_clip = HyditBertModel(dtype=dtype)
+        self.hydit_clip = HyditBertModel(dtype=dtype, model_options=model_options)
-        self.mt5xl = MT5XLModel(dtype=dtype)
+        self.mt5xl = MT5XLModel(dtype=dtype, model_options=model_options)
 
         self.dtypes = set()
         if dtype is not None:

comfy/text_encoders/long_clipl.json

@@ -0,0 +1,25 @@
+{
+  "_name_or_path": "openai/clip-vit-large-patch14",
+  "architectures": [
+    "CLIPTextModel"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 49407,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 768,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 248,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 1,
+  "projection_dim": 768,
+  "torch_dtype": "float32",
+  "transformers_version": "4.24.0",
+  "vocab_size": 49408
+}

comfy/text_encoders/long_clipl.py

@@ -0,0 +1,19 @@
+from comfy import sd1_clip
+import os
+
+class LongClipTokenizer_(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(max_length=248, embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+
+class LongClipModel_(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "long_clipl.json")
+        super().__init__(device=device, textmodel_json_config=textmodel_json_config, return_projected_pooled=False, dtype=dtype, model_options=model_options)
+
+class LongClipTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, tokenizer=LongClipTokenizer_)
+
+class LongClipModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, clip_model=LongClipModel_, **kwargs)

comfy/text_encoders/sa_t5.py

@@ -4,9 +4,9 @@ import comfy.text_encoders.t5
 import os
 
 class T5BaseModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_base.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, model_options=model_options, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True)
 
 class T5BaseTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -18,5 +18,5 @@ class SAT5Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5base", tokenizer=T5BaseTokenizer)
 
 class SAT5Model(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, **kwargs):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
-        super().__init__(device=device, dtype=dtype, name="t5base", clip_model=T5BaseModel, **kwargs)
+        super().__init__(device=device, dtype=dtype, model_options=model_options, name="t5base", clip_model=T5BaseModel, **kwargs)

comfy/text_encoders/sd2_clip.py

@@ -2,13 +2,13 @@ from comfy import sd1_clip
 import os
 
 class SD2ClipHModel(sd1_clip.SDClipModel):
-    def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None):
+    def __init__(self, arch="ViT-H-14", device="cpu", max_length=77, freeze=True, layer="penultimate", layer_idx=None, dtype=None, model_options={}):
         if layer == "penultimate":
             layer="hidden"
             layer_idx=-2
 
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd2_clip_config.json")
-        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0})
+        super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 49406, "end": 49407, "pad": 0}, return_projected_pooled=True, model_options=model_options)
 
 class SD2ClipHTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, tokenizer_path=None, embedding_directory=None, tokenizer_data={}):
@@ -19,5 +19,5 @@ class SD2Tokenizer(sd1_clip.SD1Tokenizer):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="h", tokenizer=SD2ClipHTokenizer)
 
 class SD2ClipModel(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, **kwargs):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
-        super().__init__(device=device, dtype=dtype, clip_name="h", clip_model=SD2ClipHModel, **kwargs)
+        super().__init__(device=device, dtype=dtype, model_options=model_options, clip_name="h", clip_model=SD2ClipHModel, **kwargs)

comfy/text_encoders/sd3_clip.py

@@ -8,14 +8,14 @@ import comfy.model_management
 import logging
 
 class T5XXLModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json")
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5)
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, model_options=model_options)
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=77)
+        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=77)
 
 
 class SD3Tokenizer:
@@ -38,24 +38,24 @@ class SD3Tokenizer:
         return {}
 
 class SD3ClipModel(torch.nn.Module):
-    def __init__(self, clip_l=True, clip_g=True, t5=True, dtype_t5=None, device="cpu", dtype=None):
+    def __init__(self, clip_l=True, clip_g=True, t5=True, dtype_t5=None, device="cpu", dtype=None, model_options={}):
         super().__init__()
         self.dtypes = set()
         if clip_l:
-            self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, return_projected_pooled=False)
+            self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, return_projected_pooled=False, model_options=model_options)
             self.dtypes.add(dtype)
         else:
             self.clip_l = None
 
         if clip_g:
-            self.clip_g = sdxl_clip.SDXLClipG(device=device, dtype=dtype)
+            self.clip_g = sdxl_clip.SDXLClipG(device=device, dtype=dtype, model_options=model_options)
             self.dtypes.add(dtype)
         else:
             self.clip_g = None
 
         if t5:
             dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
-            self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5)
+            self.t5xxl = T5XXLModel(device=device, dtype=dtype_t5, model_options=model_options)
             self.dtypes.add(dtype_t5)
         else:
             self.t5xxl = None
@@ -81,7 +81,7 @@ class SD3ClipModel(torch.nn.Module):
     def encode_token_weights(self, token_weight_pairs):
         token_weight_pairs_l = token_weight_pairs["l"]
         token_weight_pairs_g = token_weight_pairs["g"]
-        token_weight_pars_t5 = token_weight_pairs["t5xxl"]
+        token_weight_pairs_t5 = token_weight_pairs["t5xxl"]
         lg_out = None
         pooled = None
         out = None
@@ -108,7 +108,7 @@ class SD3ClipModel(torch.nn.Module):
             pooled = torch.cat((l_pooled, g_pooled), dim=-1)
 
         if self.t5xxl is not None:
-            t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pars_t5)
+            t5_out, t5_pooled = self.t5xxl.encode_token_weights(token_weight_pairs_t5)
             if lg_out is not None:
                 out = torch.cat([lg_out, t5_out], dim=-2)
             else:
@@ -132,6 +132,6 @@ class SD3ClipModel(torch.nn.Module):
 
 def sd3_clip(clip_l=True, clip_g=True, t5=True, dtype_t5=None):
     class SD3ClipModel_(SD3ClipModel):
-        def __init__(self, device="cpu", dtype=None):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
-            super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, dtype_t5=dtype_t5, device=device, dtype=dtype)
+            super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options)
     return SD3ClipModel_

comfy/utils.py

@@ -1,3 +1,22 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
+
 import torch
 import math
 import struct
@@ -40,9 +59,22 @@ def calculate_parameters(sd, prefix=""):
     params = 0
     for k in sd.keys():
         if k.startswith(prefix):
-            params += sd[k].nelement()
+            w = sd[k]
+            params += w.nelement()
     return params
 
+def weight_dtype(sd, prefix=""):
+    dtypes = {}
+    for k in sd.keys():
+        if k.startswith(prefix):
+            w = sd[k]
+            dtypes[w.dtype] = dtypes.get(w.dtype, 0) + 1
+
+    if len(dtypes) == 0:
+        return None
+
+    return max(dtypes, key=dtypes.get)
+
 def state_dict_key_replace(state_dict, keys_to_replace):
     for x in keys_to_replace:
         if x in state_dict:
@@ -402,6 +434,110 @@ def auraflow_to_diffusers(mmdit_config, output_prefix=""):
 
     return key_map
 
+def flux_to_diffusers(mmdit_config, output_prefix=""):
+    n_double_layers = mmdit_config.get("depth", 0)
+    n_single_layers = mmdit_config.get("depth_single_blocks", 0)
+    hidden_size = mmdit_config.get("hidden_size", 0)
+
+    key_map = {}
+    for index in range(n_double_layers):
+        prefix_from = "transformer_blocks.{}".format(index)
+        prefix_to = "{}double_blocks.{}".format(output_prefix, index)
+
+        for end in ("weight", "bias"):
+            k = "{}.attn.".format(prefix_from)
+            qkv = "{}.img_attn.qkv.{}".format(prefix_to, end)
+            key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, hidden_size))
+            key_map["{}to_k.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size))
+            key_map["{}to_v.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size))
+
+            k = "{}.attn.".format(prefix_from)
+            qkv = "{}.txt_attn.qkv.{}".format(prefix_to, end)
+            key_map["{}add_q_proj.{}".format(k, end)] = (qkv, (0, 0, hidden_size))
+            key_map["{}add_k_proj.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size))
+            key_map["{}add_v_proj.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size))
+
+        block_map = {
+                        "attn.to_out.0.weight": "img_attn.proj.weight",
+                        "attn.to_out.0.bias": "img_attn.proj.bias",
+                        "norm1.linear.weight": "img_mod.lin.weight",
+                        "norm1.linear.bias": "img_mod.lin.bias",
+                        "norm1_context.linear.weight": "txt_mod.lin.weight",
+                        "norm1_context.linear.bias": "txt_mod.lin.bias",
+                        "attn.to_add_out.weight": "txt_attn.proj.weight",
+                        "attn.to_add_out.bias": "txt_attn.proj.bias",
+                        "ff.net.0.proj.weight": "img_mlp.0.weight",
+                        "ff.net.0.proj.bias": "img_mlp.0.bias",
+                        "ff.net.2.weight": "img_mlp.2.weight",
+                        "ff.net.2.bias": "img_mlp.2.bias",
+                        "ff_context.net.0.proj.weight": "txt_mlp.0.weight",
+                        "ff_context.net.0.proj.bias": "txt_mlp.0.bias",
+                        "ff_context.net.2.weight": "txt_mlp.2.weight",
+                        "ff_context.net.2.bias": "txt_mlp.2.bias",
+                        "attn.norm_q.weight": "img_attn.norm.query_norm.scale",
+                        "attn.norm_k.weight": "img_attn.norm.key_norm.scale",
+                        "attn.norm_added_q.weight": "txt_attn.norm.query_norm.scale",
+                        "attn.norm_added_k.weight": "txt_attn.norm.key_norm.scale",
+                    }
+
+        for k in block_map:
+            key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, block_map[k])
+
+    for index in range(n_single_layers):
+        prefix_from = "single_transformer_blocks.{}".format(index)
+        prefix_to = "{}single_blocks.{}".format(output_prefix, index)
+
+        for end in ("weight", "bias"):
+            k = "{}.attn.".format(prefix_from)
+            qkv = "{}.linear1.{}".format(prefix_to, end)
+            key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, hidden_size))
+            key_map["{}to_k.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size))
+            key_map["{}to_v.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size))
+            key_map["{}.proj_mlp.{}".format(prefix_from, end)] = (qkv, (0, hidden_size * 3, hidden_size * 4))
+
+        block_map = {
+                        "norm.linear.weight": "modulation.lin.weight",
+                        "norm.linear.bias": "modulation.lin.bias",
+                        "proj_out.weight": "linear2.weight",
+                        "proj_out.bias": "linear2.bias",
+                        "attn.norm_q.weight": "norm.query_norm.scale",
+                        "attn.norm_k.weight": "norm.key_norm.scale",
+                    }
+
+        for k in block_map:
+            key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, block_map[k])
+
+    MAP_BASIC = {
+        ("final_layer.linear.bias", "proj_out.bias"),
+        ("final_layer.linear.weight", "proj_out.weight"),
+        ("img_in.bias", "x_embedder.bias"),
+        ("img_in.weight", "x_embedder.weight"),
+        ("time_in.in_layer.bias", "time_text_embed.timestep_embedder.linear_1.bias"),
+        ("time_in.in_layer.weight", "time_text_embed.timestep_embedder.linear_1.weight"),
+        ("time_in.out_layer.bias", "time_text_embed.timestep_embedder.linear_2.bias"),
+        ("time_in.out_layer.weight", "time_text_embed.timestep_embedder.linear_2.weight"),
+        ("txt_in.bias", "context_embedder.bias"),
+        ("txt_in.weight", "context_embedder.weight"),
+        ("vector_in.in_layer.bias", "time_text_embed.text_embedder.linear_1.bias"),
+        ("vector_in.in_layer.weight", "time_text_embed.text_embedder.linear_1.weight"),
+        ("vector_in.out_layer.bias", "time_text_embed.text_embedder.linear_2.bias"),
+        ("vector_in.out_layer.weight", "time_text_embed.text_embedder.linear_2.weight"),
+        ("guidance_in.in_layer.bias", "time_text_embed.guidance_embedder.linear_1.bias"),
+        ("guidance_in.in_layer.weight", "time_text_embed.guidance_embedder.linear_1.weight"),
+        ("guidance_in.out_layer.bias", "time_text_embed.guidance_embedder.linear_2.bias"),
+        ("guidance_in.out_layer.weight", "time_text_embed.guidance_embedder.linear_2.weight"),
+        ("final_layer.adaLN_modulation.1.bias", "norm_out.linear.bias", swap_scale_shift),
+        ("final_layer.adaLN_modulation.1.weight", "norm_out.linear.weight", swap_scale_shift),
+    }
+
+    for k in MAP_BASIC:
+        if len(k) > 2:
+            key_map[k[1]] = ("{}{}".format(output_prefix, k[0]), None, k[2])
+        else:
+            key_map[k[1]] = "{}{}".format(output_prefix, k[0])
+
+    return key_map
+
 def repeat_to_batch_size(tensor, batch_size, dim=0):
     if tensor.shape[dim] > batch_size:
         return tensor.narrow(dim, 0, batch_size)

comfy_execution/caching.py

@@ -0,0 +1,308 @@
+import itertools
+from typing import Sequence, Mapping
+from comfy_execution.graph import DynamicPrompt
+
+import nodes
+
+from comfy_execution.graph_utils import is_link
+
+class CacheKeySet:
+    def __init__(self, dynprompt, node_ids, is_changed_cache):
+        self.keys = {}
+        self.subcache_keys = {}
+
+    def add_keys(self, node_ids):
+        raise NotImplementedError()
+
+    def all_node_ids(self):
+        return set(self.keys.keys())
+
+    def get_used_keys(self):
+        return self.keys.values()
+
+    def get_used_subcache_keys(self):
+        return self.subcache_keys.values()
+
+    def get_data_key(self, node_id):
+        return self.keys.get(node_id, None)
+
+    def get_subcache_key(self, node_id):
+        return self.subcache_keys.get(node_id, None)
+
+class Unhashable:
+    def __init__(self):
+        self.value = float("NaN")
+
+def to_hashable(obj):
+    # So that we don't infinitely recurse since frozenset and tuples
+    # are Sequences.
+    if isinstance(obj, (int, float, str, bool, type(None))):
+        return obj
+    elif isinstance(obj, Mapping):
+        return frozenset([(to_hashable(k), to_hashable(v)) for k, v in sorted(obj.items())])
+    elif isinstance(obj, Sequence):
+        return frozenset(zip(itertools.count(), [to_hashable(i) for i in obj]))
+    else:
+        # TODO - Support other objects like tensors?
+        return Unhashable()
+
+class CacheKeySetID(CacheKeySet):
+    def __init__(self, dynprompt, node_ids, is_changed_cache):
+        super().__init__(dynprompt, node_ids, is_changed_cache)
+        self.dynprompt = dynprompt
+        self.add_keys(node_ids)
+
+    def add_keys(self, node_ids):
+        for node_id in node_ids:
+            if node_id in self.keys:
+                continue
+            if not self.dynprompt.has_node(node_id):
+                continue
+            node = self.dynprompt.get_node(node_id)
+            self.keys[node_id] = (node_id, node["class_type"])
+            self.subcache_keys[node_id] = (node_id, node["class_type"])
+
+class CacheKeySetInputSignature(CacheKeySet):
+    def __init__(self, dynprompt, node_ids, is_changed_cache):
+        super().__init__(dynprompt, node_ids, is_changed_cache)
+        self.dynprompt = dynprompt
+        self.is_changed_cache = is_changed_cache
+        self.add_keys(node_ids)
+
+    def include_node_id_in_input(self) -> bool:
+        return False
+
+    def add_keys(self, node_ids):
+        for node_id in node_ids:
+            if node_id in self.keys:
+                continue
+            if not self.dynprompt.has_node(node_id):
+                continue
+            node = self.dynprompt.get_node(node_id)
+            self.keys[node_id] = self.get_node_signature(self.dynprompt, node_id)
+            self.subcache_keys[node_id] = (node_id, node["class_type"])
+
+    def get_node_signature(self, dynprompt, node_id):
+        signature = []
+        ancestors, order_mapping = self.get_ordered_ancestry(dynprompt, node_id)
+        signature.append(self.get_immediate_node_signature(dynprompt, node_id, order_mapping))
+        for ancestor_id in ancestors:
+            signature.append(self.get_immediate_node_signature(dynprompt, ancestor_id, order_mapping))
+        return to_hashable(signature)
+
+    def get_immediate_node_signature(self, dynprompt, node_id, ancestor_order_mapping):
+        if not dynprompt.has_node(node_id):
+            # This node doesn't exist -- we can't cache it.
+            return [float("NaN")]
+        node = dynprompt.get_node(node_id)
+        class_type = node["class_type"]
+        class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
+        signature = [class_type, self.is_changed_cache.get(node_id)]
+        if self.include_node_id_in_input() or (hasattr(class_def, "NOT_IDEMPOTENT") and class_def.NOT_IDEMPOTENT):
+            signature.append(node_id)
+        inputs = node["inputs"]
+        for key in sorted(inputs.keys()):
+            if is_link(inputs[key]):
+                (ancestor_id, ancestor_socket) = inputs[key]
+                ancestor_index = ancestor_order_mapping[ancestor_id]
+                signature.append((key,("ANCESTOR", ancestor_index, ancestor_socket)))
+            else:
+                signature.append((key, inputs[key]))
+        return signature
+
+    # This function returns a list of all ancestors of the given node. The order of the list is
+    # deterministic based on which specific inputs the ancestor is connected by.
+    def get_ordered_ancestry(self, dynprompt, node_id):
+        ancestors = []
+        order_mapping = {}
+        self.get_ordered_ancestry_internal(dynprompt, node_id, ancestors, order_mapping)
+        return ancestors, order_mapping
+
+    def get_ordered_ancestry_internal(self, dynprompt, node_id, ancestors, order_mapping):
+        if not dynprompt.has_node(node_id):
+            return
+        inputs = dynprompt.get_node(node_id)["inputs"]
+        input_keys = sorted(inputs.keys())
+        for key in input_keys:
+            if is_link(inputs[key]):
+                ancestor_id = inputs[key][0]
+                if ancestor_id not in order_mapping:
+                    ancestors.append(ancestor_id)
+                    order_mapping[ancestor_id] = len(ancestors) - 1
+                    self.get_ordered_ancestry_internal(dynprompt, ancestor_id, ancestors, order_mapping)
+
+class BasicCache:
+    def __init__(self, key_class):
+        self.key_class = key_class
+        self.initialized = False
+        self.dynprompt: DynamicPrompt
+        self.cache_key_set: CacheKeySet
+        self.cache = {}
+        self.subcaches = {}
+
+    def set_prompt(self, dynprompt, node_ids, is_changed_cache):
+        self.dynprompt = dynprompt
+        self.cache_key_set = self.key_class(dynprompt, node_ids, is_changed_cache)
+        self.is_changed_cache = is_changed_cache
+        self.initialized = True
+
+    def all_node_ids(self):
+        assert self.initialized
+        node_ids = self.cache_key_set.all_node_ids()
+        for subcache in self.subcaches.values():
+            node_ids = node_ids.union(subcache.all_node_ids())
+        return node_ids
+
+    def _clean_cache(self):
+        preserve_keys = set(self.cache_key_set.get_used_keys())
+        to_remove = []
+        for key in self.cache:
+            if key not in preserve_keys:
+                to_remove.append(key)
+        for key in to_remove:
+            del self.cache[key]
+
+    def _clean_subcaches(self):
+        preserve_subcaches = set(self.cache_key_set.get_used_subcache_keys())
+
+        to_remove = []
+        for key in self.subcaches:
+            if key not in preserve_subcaches:
+                to_remove.append(key)
+        for key in to_remove:
+            del self.subcaches[key]
+
+    def clean_unused(self):
+        assert self.initialized
+        self._clean_cache()
+        self._clean_subcaches()
+
+    def _set_immediate(self, node_id, value):
+        assert self.initialized
+        cache_key = self.cache_key_set.get_data_key(node_id)
+        self.cache[cache_key] = value
+
+    def _get_immediate(self, node_id):
+        if not self.initialized:
+            return None
+        cache_key = self.cache_key_set.get_data_key(node_id)
+        if cache_key in self.cache:
+            return self.cache[cache_key]
+        else:
+            return None
+
+    def _ensure_subcache(self, node_id, children_ids):
+        subcache_key = self.cache_key_set.get_subcache_key(node_id)
+        subcache = self.subcaches.get(subcache_key, None)
+        if subcache is None:
+            subcache = BasicCache(self.key_class)
+            self.subcaches[subcache_key] = subcache
+        subcache.set_prompt(self.dynprompt, children_ids, self.is_changed_cache)
+        return subcache
+
+    def _get_subcache(self, node_id):
+        assert self.initialized
+        subcache_key = self.cache_key_set.get_subcache_key(node_id)
+        if subcache_key in self.subcaches:
+            return self.subcaches[subcache_key]
+        else:
+            return None
+
+    def recursive_debug_dump(self):
+        result = []
+        for key in self.cache:
+            result.append({"key": key, "value": self.cache[key]})
+        for key in self.subcaches:
+            result.append({"subcache_key": key, "subcache": self.subcaches[key].recursive_debug_dump()})
+        return result
+
+class HierarchicalCache(BasicCache):
+    def __init__(self, key_class):
+        super().__init__(key_class)
+
+    def _get_cache_for(self, node_id):
+        assert self.dynprompt is not None
+        parent_id = self.dynprompt.get_parent_node_id(node_id)
+        if parent_id is None:
+            return self
+
+        hierarchy = []
+        while parent_id is not None:
+            hierarchy.append(parent_id)
+            parent_id = self.dynprompt.get_parent_node_id(parent_id)
+
+        cache = self
+        for parent_id in reversed(hierarchy):
+            cache = cache._get_subcache(parent_id)
+            if cache is None:
+                return None
+        return cache
+
+    def get(self, node_id):
+        cache = self._get_cache_for(node_id)
+        if cache is None:
+            return None
+        return cache._get_immediate(node_id)
+
+    def set(self, node_id, value):
+        cache = self._get_cache_for(node_id)
+        assert cache is not None
+        cache._set_immediate(node_id, value)
+
+    def ensure_subcache_for(self, node_id, children_ids):
+        cache = self._get_cache_for(node_id)
+        assert cache is not None
+        return cache._ensure_subcache(node_id, children_ids)
+
+class LRUCache(BasicCache):
+    def __init__(self, key_class, max_size=100):
+        super().__init__(key_class)
+        self.max_size = max_size
+        self.min_generation = 0
+        self.generation = 0
+        self.used_generation = {}
+        self.children = {}
+
+    def set_prompt(self, dynprompt, node_ids, is_changed_cache):
+        super().set_prompt(dynprompt, node_ids, is_changed_cache)
+        self.generation += 1
+        for node_id in node_ids:
+            self._mark_used(node_id)
+
+    def clean_unused(self):
+        while len(self.cache) > self.max_size and self.min_generation < self.generation:
+            self.min_generation += 1
+            to_remove = [key for key in self.cache if self.used_generation[key] < self.min_generation]
+            for key in to_remove:
+                del self.cache[key]
+                del self.used_generation[key]
+                if key in self.children:
+                    del self.children[key]
+        self._clean_subcaches()
+
+    def get(self, node_id):
+        self._mark_used(node_id)
+        return self._get_immediate(node_id)
+
+    def _mark_used(self, node_id):
+        cache_key = self.cache_key_set.get_data_key(node_id)
+        if cache_key is not None:
+            self.used_generation[cache_key] = self.generation
+
+    def set(self, node_id, value):
+        self._mark_used(node_id)
+        return self._set_immediate(node_id, value)
+
+    def ensure_subcache_for(self, node_id, children_ids):
+        # Just uses subcaches for tracking 'live' nodes
+        super()._ensure_subcache(node_id, children_ids)
+
+        self.cache_key_set.add_keys(children_ids)
+        self._mark_used(node_id)
+        cache_key = self.cache_key_set.get_data_key(node_id)
+        self.children[cache_key] = []
+        for child_id in children_ids:
+            self._mark_used(child_id)
+            self.children[cache_key].append(self.cache_key_set.get_data_key(child_id))
+        return self
+

comfy_execution/graph.py

@@ -0,0 +1,259 @@
+import nodes
+
+from comfy_execution.graph_utils import is_link
+
+class DependencyCycleError(Exception):
+    pass
+
+class NodeInputError(Exception):
+    pass
+
+class NodeNotFoundError(Exception):
+    pass
+
+class DynamicPrompt:
+    def __init__(self, original_prompt):
+        # The original prompt provided by the user
+        self.original_prompt = original_prompt
+        # Any extra pieces of the graph created during execution
+        self.ephemeral_prompt = {}
+        self.ephemeral_parents = {}
+        self.ephemeral_display = {}
+
+    def get_node(self, node_id):
+        if node_id in self.ephemeral_prompt:
+            return self.ephemeral_prompt[node_id]
+        if node_id in self.original_prompt:
+            return self.original_prompt[node_id]
+        raise NodeNotFoundError(f"Node {node_id} not found")
+
+    def has_node(self, node_id):
+        return node_id in self.original_prompt or node_id in self.ephemeral_prompt
+
+    def add_ephemeral_node(self, node_id, node_info, parent_id, display_id):
+        self.ephemeral_prompt[node_id] = node_info
+        self.ephemeral_parents[node_id] = parent_id
+        self.ephemeral_display[node_id] = display_id
+
+    def get_real_node_id(self, node_id):
+        while node_id in self.ephemeral_parents:
+            node_id = self.ephemeral_parents[node_id]
+        return node_id
+
+    def get_parent_node_id(self, node_id):
+        return self.ephemeral_parents.get(node_id, None)
+
+    def get_display_node_id(self, node_id):
+        while node_id in self.ephemeral_display:
+            node_id = self.ephemeral_display[node_id]
+        return node_id
+
+    def all_node_ids(self):
+        return set(self.original_prompt.keys()).union(set(self.ephemeral_prompt.keys()))
+
+    def get_original_prompt(self):
+        return self.original_prompt
+
+def get_input_info(class_def, input_name):
+    valid_inputs = class_def.INPUT_TYPES()
+    input_info = None
+    input_category = None
+    if "required" in valid_inputs and input_name in valid_inputs["required"]:
+        input_category = "required"
+        input_info = valid_inputs["required"][input_name]
+    elif "optional" in valid_inputs and input_name in valid_inputs["optional"]:
+        input_category = "optional"
+        input_info = valid_inputs["optional"][input_name]
+    elif "hidden" in valid_inputs and input_name in valid_inputs["hidden"]:
+        input_category = "hidden"
+        input_info = valid_inputs["hidden"][input_name]
+    if input_info is None:
+        return None, None, None
+    input_type = input_info[0]
+    if len(input_info) > 1:
+        extra_info = input_info[1]
+    else:
+        extra_info = {}
+    return input_type, input_category, extra_info
+
+class TopologicalSort:
+    def __init__(self, dynprompt):
+        self.dynprompt = dynprompt
+        self.pendingNodes = {}
+        self.blockCount = {} # Number of nodes this node is directly blocked by
+        self.blocking = {} # Which nodes are blocked by this node
+
+    def get_input_info(self, unique_id, input_name):
+        class_type = self.dynprompt.get_node(unique_id)["class_type"]
+        class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
+        return get_input_info(class_def, input_name)
+
+    def make_input_strong_link(self, to_node_id, to_input):
+        inputs = self.dynprompt.get_node(to_node_id)["inputs"]
+        if to_input not in inputs:
+            raise NodeInputError(f"Node {to_node_id} says it needs input {to_input}, but there is no input to that node at all")
+        value = inputs[to_input]
+        if not is_link(value):
+            raise NodeInputError(f"Node {to_node_id} says it needs input {to_input}, but that value is a constant")
+        from_node_id, from_socket = value
+        self.add_strong_link(from_node_id, from_socket, to_node_id)
+
+    def add_strong_link(self, from_node_id, from_socket, to_node_id):
+        self.add_node(from_node_id)
+        if to_node_id not in self.blocking[from_node_id]:
+            self.blocking[from_node_id][to_node_id] = {}
+            self.blockCount[to_node_id] += 1
+        self.blocking[from_node_id][to_node_id][from_socket] = True
+
+    def add_node(self, unique_id, include_lazy=False, subgraph_nodes=None):
+        if unique_id in self.pendingNodes:
+            return
+        self.pendingNodes[unique_id] = True
+        self.blockCount[unique_id] = 0
+        self.blocking[unique_id] = {}
+
+        inputs = self.dynprompt.get_node(unique_id)["inputs"]
+        for input_name in inputs:
+            value = inputs[input_name]
+            if is_link(value):
+                from_node_id, from_socket = value
+                if subgraph_nodes is not None and from_node_id not in subgraph_nodes:
+                    continue
+                input_type, input_category, input_info = self.get_input_info(unique_id, input_name)
+                is_lazy = input_info is not None and "lazy" in input_info and input_info["lazy"]
+                if include_lazy or not is_lazy:
+                    self.add_strong_link(from_node_id, from_socket, unique_id)
+
+    def get_ready_nodes(self):
+        return [node_id for node_id in self.pendingNodes if self.blockCount[node_id] == 0]
+
+    def pop_node(self, unique_id):
+        del self.pendingNodes[unique_id]
+        for blocked_node_id in self.blocking[unique_id]:
+            self.blockCount[blocked_node_id] -= 1
+        del self.blocking[unique_id]
+
+    def is_empty(self):
+        return len(self.pendingNodes) == 0
+
+class ExecutionList(TopologicalSort):
+    """
+    ExecutionList implements a topological dissolve of the graph. After a node is staged for execution,
+    it can still be returned to the graph after having further dependencies added.
+    """
+    def __init__(self, dynprompt, output_cache):
+        super().__init__(dynprompt)
+        self.output_cache = output_cache
+        self.staged_node_id = None
+
+    def add_strong_link(self, from_node_id, from_socket, to_node_id):
+        if self.output_cache.get(from_node_id) is not None:
+            # Nothing to do
+            return
+        super().add_strong_link(from_node_id, from_socket, to_node_id)
+
+    def stage_node_execution(self):
+        assert self.staged_node_id is None
+        if self.is_empty():
+            return None, None, None
+        available = self.get_ready_nodes()
+        if len(available) == 0:
+            cycled_nodes = self.get_nodes_in_cycle()
+            # Because cycles composed entirely of static nodes are caught during initial validation,
+            # we will 'blame' the first node in the cycle that is not a static node.
+            blamed_node = cycled_nodes[0]
+            for node_id in cycled_nodes:
+                display_node_id = self.dynprompt.get_display_node_id(node_id)
+                if display_node_id != node_id:
+                    blamed_node = display_node_id
+                    break
+            ex = DependencyCycleError("Dependency cycle detected")
+            error_details = {
+                "node_id": blamed_node,
+                "exception_message": str(ex),
+                "exception_type": "graph.DependencyCycleError",
+                "traceback": [],
+                "current_inputs": []
+            }
+            return None, error_details, ex
+
+        self.staged_node_id = self.ux_friendly_pick_node(available)
+        return self.staged_node_id, None, None
+
+    def ux_friendly_pick_node(self, node_list):
+        # If an output node is available, do that first.
+        # Technically this has no effect on the overall length of execution, but it feels better as a user
+        # for a PreviewImage to display a result as soon as it can
+        # Some other heuristics could probably be used here to improve the UX further.
+        def is_output(node_id):
+            class_type = self.dynprompt.get_node(node_id)["class_type"]
+            class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
+            if hasattr(class_def, 'OUTPUT_NODE') and class_def.OUTPUT_NODE == True:
+                return True
+            return False
+
+        for node_id in node_list:
+            if is_output(node_id):
+                return node_id
+
+        #This should handle the VAEDecode -> preview case
+        for node_id in node_list:
+            for blocked_node_id in self.blocking[node_id]:
+                if is_output(blocked_node_id):
+                    return node_id
+
+        #This should handle the VAELoader -> VAEDecode -> preview case
+        for node_id in node_list:
+            for blocked_node_id in self.blocking[node_id]:
+                for blocked_node_id1 in self.blocking[blocked_node_id]:
+                    if is_output(blocked_node_id1):
+                        return node_id
+
+        #TODO: this function should be improved
+        return node_list[0]
+
+    def unstage_node_execution(self):
+        assert self.staged_node_id is not None
+        self.staged_node_id = None
+
+    def complete_node_execution(self):
+        node_id = self.staged_node_id
+        self.pop_node(node_id)
+        self.staged_node_id = None
+
+    def get_nodes_in_cycle(self):
+        # We'll dissolve the graph in reverse topological order to leave only the nodes in the cycle.
+        # We're skipping some of the performance optimizations from the original TopologicalSort to keep
+        # the code simple (and because having a cycle in the first place is a catastrophic error)
+        blocked_by = { node_id: {} for node_id in self.pendingNodes }
+        for from_node_id in self.blocking:
+            for to_node_id in self.blocking[from_node_id]:
+                if True in self.blocking[from_node_id][to_node_id].values():
+                    blocked_by[to_node_id][from_node_id] = True
+        to_remove = [node_id for node_id in blocked_by if len(blocked_by[node_id]) == 0]
+        while len(to_remove) > 0:
+            for node_id in to_remove:
+                for to_node_id in blocked_by:
+                    if node_id in blocked_by[to_node_id]:
+                        del blocked_by[to_node_id][node_id]
+                del blocked_by[node_id]
+            to_remove = [node_id for node_id in blocked_by if len(blocked_by[node_id]) == 0]
+        return list(blocked_by.keys())
+
+class ExecutionBlocker:
+    """
+    Return this from a node and any users will be blocked with the given error message.
+    If the message is None, execution will be blocked silently instead.
+    Generally, you should avoid using this functionality unless absolutely necessary. Whenever it's
+    possible, a lazy input will be more efficient and have a better user experience.
+    This functionality is useful in two cases:
+    1. You want to conditionally prevent an output node from executing. (Particularly a built-in node
+       like SaveImage. For your own output nodes, I would recommend just adding a BOOL input and using
+       lazy evaluation to let it conditionally disable itself.)
+    2. You have a node with multiple possible outputs, some of which are invalid and should not be used.
+       (I would recommend not making nodes like this in the future -- instead, make multiple nodes with
+       different outputs. Unfortunately, there are several popular existing nodes using this pattern.)
+    """
+    def __init__(self, message):
+        self.message = message
+

comfy_execution/graph_utils.py

@@ -0,0 +1,139 @@
+def is_link(obj):
+    if not isinstance(obj, list):
+        return False
+    if len(obj) != 2:
+        return False
+    if not isinstance(obj[0], str):
+        return False
+    if not isinstance(obj[1], int) and not isinstance(obj[1], float):
+        return False
+    return True
+
+# The GraphBuilder is just a utility class that outputs graphs in the form expected by the ComfyUI back-end
+class GraphBuilder:
+    _default_prefix_root = ""
+    _default_prefix_call_index = 0
+    _default_prefix_graph_index = 0
+
+    def __init__(self, prefix = None):
+        if prefix is None:
+            self.prefix = GraphBuilder.alloc_prefix()
+        else:
+            self.prefix = prefix
+        self.nodes = {}
+        self.id_gen = 1
+
+    @classmethod
+    def set_default_prefix(cls, prefix_root, call_index, graph_index = 0):
+        cls._default_prefix_root = prefix_root
+        cls._default_prefix_call_index = call_index
+        cls._default_prefix_graph_index = graph_index
+
+    @classmethod
+    def alloc_prefix(cls, root=None, call_index=None, graph_index=None):
+        if root is None:
+            root = GraphBuilder._default_prefix_root
+        if call_index is None:
+            call_index = GraphBuilder._default_prefix_call_index
+        if graph_index is None:
+            graph_index = GraphBuilder._default_prefix_graph_index
+        result = f"{root}.{call_index}.{graph_index}."
+        GraphBuilder._default_prefix_graph_index += 1
+        return result
+
+    def node(self, class_type, id=None, **kwargs):
+        if id is None:
+            id = str(self.id_gen)
+            self.id_gen += 1
+        id = self.prefix + id
+        if id in self.nodes:
+            return self.nodes[id]
+
+        node = Node(id, class_type, kwargs)
+        self.nodes[id] = node
+        return node
+
+    def lookup_node(self, id):
+        id = self.prefix + id
+        return self.nodes.get(id)
+
+    def finalize(self):
+        output = {}
+        for node_id, node in self.nodes.items():
+            output[node_id] = node.serialize()
+        return output
+
+    def replace_node_output(self, node_id, index, new_value):
+        node_id = self.prefix + node_id
+        to_remove = []
+        for node in self.nodes.values():
+            for key, value in node.inputs.items():
+                if is_link(value) and value[0] == node_id and value[1] == index:
+                    if new_value is None:
+                        to_remove.append((node, key))
+                    else:
+                        node.inputs[key] = new_value
+        for node, key in to_remove:
+            del node.inputs[key]
+
+    def remove_node(self, id):
+        id = self.prefix + id
+        del self.nodes[id]
+
+class Node:
+    def __init__(self, id, class_type, inputs):
+        self.id = id
+        self.class_type = class_type
+        self.inputs = inputs
+        self.override_display_id = None
+
+    def out(self, index):
+        return [self.id, index]
+
+    def set_input(self, key, value):
+        if value is None:
+            if key in self.inputs:
+                del self.inputs[key]
+        else:
+            self.inputs[key] = value
+
+    def get_input(self, key):
+        return self.inputs.get(key)
+
+    def set_override_display_id(self, override_display_id):
+        self.override_display_id = override_display_id
+
+    def serialize(self):
+        serialized = {
+            "class_type": self.class_type,
+            "inputs": self.inputs
+        }
+        if self.override_display_id is not None:
+            serialized["override_display_id"] = self.override_display_id
+        return serialized
+
+def add_graph_prefix(graph, outputs, prefix):
+    # Change the node IDs and any internal links
+    new_graph = {}
+    for node_id, node_info in graph.items():
+        # Make sure the added nodes have unique IDs
+        new_node_id = prefix + node_id
+        new_node = { "class_type": node_info["class_type"], "inputs": {} }
+        for input_name, input_value in node_info.get("inputs", {}).items():
+            if is_link(input_value):
+                new_node["inputs"][input_name] = [prefix + input_value[0], input_value[1]]
+            else:
+                new_node["inputs"][input_name] = input_value
+        new_graph[new_node_id] = new_node
+
+    # Change the node IDs in the outputs
+    new_outputs = []
+    for n in range(len(outputs)):
+        output = outputs[n]
+        if is_link(output):
+            new_outputs.append([prefix + output[0], output[1]])
+        else:
+            new_outputs.append(output)
+
+    return new_graph, tuple(new_outputs)
+

comfy_extras/nodes_hunyuan.py

@@ -19,6 +19,7 @@ class CLIPTextEncodeHunyuanDiT:
         cond = output.pop("cond")
         return ([[cond, output]], )
 
+
 NODE_CLASS_MAPPINGS = {
     "CLIPTextEncodeHunyuanDiT": CLIPTextEncodeHunyuanDiT,
 }

comfy_extras/nodes_model_advanced.py

@@ -2,6 +2,7 @@ import folder_paths
 import comfy.sd
 import comfy.model_sampling
 import comfy.latent_formats
+import nodes
 import torch
 
 class LCM(comfy.model_sampling.EPS):
@@ -170,6 +171,42 @@ class ModelSamplingAuraFlow(ModelSamplingSD3):
     def patch_aura(self, model, shift):
         return self.patch(model, shift, multiplier=1.0)
 
+class ModelSamplingFlux:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "max_shift": ("FLOAT", {"default": 1.15, "min": 0.0, "max": 100.0, "step":0.01}),
+                              "base_shift": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 100.0, "step":0.01}),
+                              "width": ("INT", {"default": 1024, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}),
+                              "height": ("INT", {"default": 1024, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}),
+                              }}
+
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "patch"
+
+    CATEGORY = "advanced/model"
+
+    def patch(self, model, max_shift, base_shift, width, height):
+        m = model.clone()
+
+        x1 = 256
+        x2 = 4096
+        mm = (max_shift - base_shift) / (x2 - x1)
+        b = base_shift - mm * x1
+        shift = (width * height / (8 * 8 * 2 * 2)) * mm + b
+
+        sampling_base = comfy.model_sampling.ModelSamplingFlux
+        sampling_type = comfy.model_sampling.CONST
+
+        class ModelSamplingAdvanced(sampling_base, sampling_type):
+            pass
+
+        model_sampling = ModelSamplingAdvanced(model.model.model_config)
+        model_sampling.set_parameters(shift=shift)
+        m.add_object_patch("model_sampling", model_sampling)
+        return (m, )
+
+
 class ModelSamplingContinuousEDM:
     @classmethod
     def INPUT_TYPES(s):
@@ -284,5 +321,6 @@ NODE_CLASS_MAPPINGS = {
     "ModelSamplingStableCascade": ModelSamplingStableCascade,
     "ModelSamplingSD3": ModelSamplingSD3,
     "ModelSamplingAuraFlow": ModelSamplingAuraFlow,
+    "ModelSamplingFlux": ModelSamplingFlux,
     "RescaleCFG": RescaleCFG,
 }

comfy_extras/nodes_model_merging.py

@@ -264,6 +264,7 @@ class CLIPSave:
 
         metadata = {}
         if not args.disable_metadata:
+            metadata["format"] = "pt"
             metadata["prompt"] = prompt_info
             if extra_pnginfo is not None:
                 for x in extra_pnginfo:
@@ -332,6 +333,25 @@ class VAESave:
         comfy.utils.save_torch_file(vae.get_sd(), output_checkpoint, metadata=metadata)
         return {}
 
+class ModelSave:
+    def __init__(self):
+        self.output_dir = folder_paths.get_output_directory()
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "filename_prefix": ("STRING", {"default": "diffusion_models/ComfyUI"}),},
+                "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"},}
+    RETURN_TYPES = ()
+    FUNCTION = "save"
+    OUTPUT_NODE = True
+
+    CATEGORY = "advanced/model_merging"
+
+    def save(self, model, filename_prefix, prompt=None, extra_pnginfo=None):
+        save_checkpoint(model, filename_prefix=filename_prefix, output_dir=self.output_dir, prompt=prompt, extra_pnginfo=extra_pnginfo)
+        return {}
+
 NODE_CLASS_MAPPINGS = {
     "ModelMergeSimple": ModelMergeSimple,
     "ModelMergeBlocks": ModelMergeBlocks,
@@ -343,4 +363,9 @@ NODE_CLASS_MAPPINGS = {
     "CLIPMergeAdd": CLIPAdd,
     "CLIPSave": CLIPSave,
     "VAESave": VAESave,
+    "ModelSave": ModelSave,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "CheckpointSave": "Save Checkpoint",
 }

comfy_extras/nodes_model_merging_model_specific.py

@@ -75,9 +75,36 @@ class ModelMergeSD3_2B(comfy_extras.nodes_model_merging.ModelMergeBlocks):
 
         return {"required": arg_dict}
 
+class ModelMergeFlux1(comfy_extras.nodes_model_merging.ModelMergeBlocks):
+    CATEGORY = "advanced/model_merging/model_specific"
+
+    @classmethod
+    def INPUT_TYPES(s):
+        arg_dict = { "model1": ("MODEL",),
+                              "model2": ("MODEL",)}
+
+        argument = ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01})
+
+        arg_dict["img_in."] = argument
+        arg_dict["time_in."] = argument
+        arg_dict["guidance_in"] = argument
+        arg_dict["vector_in."] = argument
+        arg_dict["txt_in."] = argument
+
+        for i in range(19):
+            arg_dict["double_blocks.{}.".format(i)] = argument
+
+        for i in range(38):
+            arg_dict["single_blocks.{}.".format(i)] = argument
+
+        arg_dict["final_layer."] = argument
+
+        return {"required": arg_dict}
+
 NODE_CLASS_MAPPINGS = {
     "ModelMergeSD1": ModelMergeSD1,
     "ModelMergeSD2": ModelMergeSD1, #SD1 and SD2 have the same blocks
     "ModelMergeSDXL": ModelMergeSDXL,
     "ModelMergeSD3_2B": ModelMergeSD3_2B,
+    "ModelMergeFlux1": ModelMergeFlux1,
 }

comfy_extras/nodes_sd3.py

@@ -27,8 +27,8 @@ class EmptySD3LatentImage:
 
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": { "width": ("INT", {"default": 1024, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}),
+        return {"required": { "width": ("INT", {"default": 1024, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                              "height": ("INT", {"default": 1024, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 8}),
+                              "height": ("INT", {"default": 1024, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
                               "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096})}}
     RETURN_TYPES = ("LATENT",)
     FUNCTION = "generate"
@@ -100,3 +100,8 @@ NODE_CLASS_MAPPINGS = {
     "CLIPTextEncodeSD3": CLIPTextEncodeSD3,
     "ControlNetApplySD3": ControlNetApplySD3,
 }
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    # Sampling
+    "ControlNetApplySD3": "ControlNetApply SD3 and HunyuanDiT",
+}

custom_nodes/example_node.py.example

@@ -4,14 +4,14 @@ class Example:
 
     Class methods
     -------------
-    INPUT_TYPES (dict): 
+    INPUT_TYPES (dict):
         Tell the main program input parameters of nodes.
     IS_CHANGED:
         optional method to control when the node is re executed.
 
     Attributes
     ----------
-    RETURN_TYPES (`tuple`): 
+    RETURN_TYPES (`tuple`):
         The type of each element in the output tuple.
     RETURN_NAMES (`tuple`):
         Optional: The name of each output in the output tuple.
@@ -23,13 +23,19 @@ class Example:
         Assumed to be False if not present.
     CATEGORY (`str`):
         The category the node should appear in the UI.
+    DEPRECATED (`bool`):
+        Indicates whether the node is deprecated. Deprecated nodes are hidden by default in the UI, but remain
+        functional in existing workflows that use them.
+    EXPERIMENTAL (`bool`):
+        Indicates whether the node is experimental. Experimental nodes are marked as such in the UI and may be subject to
+        significant changes or removal in future versions. Use with caution in production workflows.
     execute(s) -> tuple || None:
         The entry point method. The name of this method must be the same as the value of property `FUNCTION`.
         For example, if `FUNCTION = "execute"` then this method's name must be `execute`, if `FUNCTION = "foo"` then it must be `foo`.
     """
     def __init__(self):
         pass
-    
+
     @classmethod
     def INPUT_TYPES(s):
         """
@@ -54,7 +60,8 @@ class Example:
                     "min": 0, #Minimum value
                     "max": 4096, #Maximum value
                     "step": 64, #Slider's step
-                    "display": "number" # Cosmetic only: display as "number" or "slider"
+                    "display": "number", # Cosmetic only: display as "number" or "slider"
+                    "lazy": True # Will only be evaluated if check_lazy_status requires it
                 }),
                 "float_field": ("FLOAT", {
                     "default": 1.0,
@@ -62,11 +69,14 @@ class Example:
                     "max": 10.0,
                     "step": 0.01,
                     "round": 0.001, #The value representing the precision to round to, will be set to the step value by default. Can be set to False to disable rounding.
-                    "display": "number"}),
+                    "display": "number",
+                    "lazy": True
+                }),
                 "print_to_screen": (["enable", "disable"],),
                 "string_field": ("STRING", {
                     "multiline": False, #True if you want the field to look like the one on the ClipTextEncode node
-                    "default": "Hello World!"
+                    "default": "Hello World!",
+                    "lazy": True
                 }),
             },
         }
@@ -80,6 +90,23 @@ class Example:
 
     CATEGORY = "Example"
 
+    def check_lazy_status(self, image, string_field, int_field, float_field, print_to_screen):
+        """
+            Return a list of input names that need to be evaluated.
+
+            This function will be called if there are any lazy inputs which have not yet been
+            evaluated. As long as you return at least one field which has not yet been evaluated
+            (and more exist), this function will be called again once the value of the requested
+            field is available.
+
+            Any evaluated inputs will be passed as arguments to this function. Any unevaluated
+            inputs will have the value None.
+        """
+        if print_to_screen == "enable":
+            return ["int_field", "float_field", "string_field"]
+        else:
+            return []
+
     def test(self, image, string_field, int_field, float_field, print_to_screen):
         if print_to_screen == "enable":
             print(f"""Your input contains:

execution.py

@@ -5,6 +5,7 @@ import threading
 import heapq
 import time
 import traceback
+from enum import Enum
 import inspect
 from typing import List, Literal, NamedTuple, Optional
 
@@ -12,102 +13,219 @@ import torch
 import nodes
 
 import comfy.model_management
+from comfy_execution.graph import get_input_info, ExecutionList, DynamicPrompt, ExecutionBlocker
+from comfy_execution.graph_utils import is_link, GraphBuilder
+from comfy_execution.caching import HierarchicalCache, LRUCache, CacheKeySetInputSignature, CacheKeySetID
+from comfy.cli_args import args
 
-def get_input_data(inputs, class_def, unique_id, outputs={}, prompt={}, extra_data={}):
+class ExecutionResult(Enum):
+    SUCCESS = 0
+    FAILURE = 1
+    PENDING = 2
+
+class DuplicateNodeError(Exception):
+    pass
+
+class IsChangedCache:
+    def __init__(self, dynprompt, outputs_cache):
+        self.dynprompt = dynprompt
+        self.outputs_cache = outputs_cache
+        self.is_changed = {}
+
+    def get(self, node_id):
+        if node_id in self.is_changed:
+            return self.is_changed[node_id]
+
+        node = self.dynprompt.get_node(node_id)
+        class_type = node["class_type"]
+        class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
+        if not hasattr(class_def, "IS_CHANGED"):
+            self.is_changed[node_id] = False
+            return self.is_changed[node_id]
+
+        if "is_changed" in node:
+            self.is_changed[node_id] = node["is_changed"]
+            return self.is_changed[node_id]
+
+        # Intentionally do not use cached outputs here. We only want constants in IS_CHANGED
+        input_data_all, _ = get_input_data(node["inputs"], class_def, node_id, None)
+        try:
+            is_changed = _map_node_over_list(class_def, input_data_all, "IS_CHANGED")
+            node["is_changed"] = [None if isinstance(x, ExecutionBlocker) else x for x in is_changed]
+        except Exception as e:
+            logging.warning("WARNING: {}".format(e))
+            node["is_changed"] = float("NaN")
+        finally:
+            self.is_changed[node_id] = node["is_changed"]
+        return self.is_changed[node_id]
+
+class CacheSet:
+    def __init__(self, lru_size=None):
+        if lru_size is None or lru_size == 0:
+            self.init_classic_cache() 
+        else:
+            self.init_lru_cache(lru_size)
+        self.all = [self.outputs, self.ui, self.objects]
+
+    # Useful for those with ample RAM/VRAM -- allows experimenting without
+    # blowing away the cache every time
+    def init_lru_cache(self, cache_size):
+        self.outputs = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
+        self.ui = LRUCache(CacheKeySetInputSignature, max_size=cache_size)
+        self.objects = HierarchicalCache(CacheKeySetID)
+
+    # Performs like the old cache -- dump data ASAP
+    def init_classic_cache(self):
+        self.outputs = HierarchicalCache(CacheKeySetInputSignature)
+        self.ui = HierarchicalCache(CacheKeySetInputSignature)
+        self.objects = HierarchicalCache(CacheKeySetID)
+
+    def recursive_debug_dump(self):
+        result = {
+            "outputs": self.outputs.recursive_debug_dump(),
+            "ui": self.ui.recursive_debug_dump(),
+        }
+        return result
+
+def get_input_data(inputs, class_def, unique_id, outputs=None, dynprompt=None, extra_data={}):
     valid_inputs = class_def.INPUT_TYPES()
     input_data_all = {}
+    missing_keys = {}
     for x in inputs:
         input_data = inputs[x]
-        if isinstance(input_data, list):
+        input_type, input_category, input_info = get_input_info(class_def, x)
+        def mark_missing():
+            missing_keys[x] = True
+            input_data_all[x] = (None,)
+        if is_link(input_data) and (not input_info or not input_info.get("rawLink", False)):
             input_unique_id = input_data[0]
             output_index = input_data[1]
-            if input_unique_id not in outputs:
+            if outputs is None:
-                input_data_all[x] = (None,)
+                mark_missing()
+                continue # This might be a lazily-evaluated input
+            cached_output = outputs.get(input_unique_id)
+            if cached_output is None:
+                mark_missing()
                 continue
-            obj = outputs[input_unique_id][output_index]
+            if output_index >= len(cached_output):
+                mark_missing()
+                continue
+            obj = cached_output[output_index]
             input_data_all[x] = obj
-        else:
+        elif input_category is not None:
-            if ("required" in valid_inputs and x in valid_inputs["required"]) or ("optional" in valid_inputs and x in valid_inputs["optional"]):
+            input_data_all[x] = [input_data]
-                input_data_all[x] = [input_data]
 
     if "hidden" in valid_inputs:
         h = valid_inputs["hidden"]
         for x in h:
             if h[x] == "PROMPT":
-                input_data_all[x] = [prompt]
+                input_data_all[x] = [dynprompt.get_original_prompt() if dynprompt is not None else {}]
+            if h[x] == "DYNPROMPT":
+                input_data_all[x] = [dynprompt]
             if h[x] == "EXTRA_PNGINFO":
                 input_data_all[x] = [extra_data.get('extra_pnginfo', None)]
             if h[x] == "UNIQUE_ID":
                 input_data_all[x] = [unique_id]
-    return input_data_all
+    return input_data_all, missing_keys
 
-def map_node_over_list(obj, input_data_all, func, allow_interrupt=False):
+map_node_over_list = None #Don't hook this please
+
+def _map_node_over_list(obj, input_data_all, func, allow_interrupt=False, execution_block_cb=None, pre_execute_cb=None):
     # check if node wants the lists
-    input_is_list = False
+    input_is_list = getattr(obj, "INPUT_IS_LIST", False)
-    if hasattr(obj, "INPUT_IS_LIST"):
-        input_is_list = obj.INPUT_IS_LIST
 
     if len(input_data_all) == 0:
         max_len_input = 0
     else:
-        max_len_input = max([len(x) for x in input_data_all.values()])
+        max_len_input = max(len(x) for x in input_data_all.values())
      
     # get a slice of inputs, repeat last input when list isn't long enough
     def slice_dict(d, i):
-        d_new = dict()
+        return {k: v[i if len(v) > i else -1] for k, v in d.items()}
-        for k,v in d.items():
-            d_new[k] = v[i if len(v) > i else -1]
-        return d_new
     
     results = []
+    def process_inputs(inputs, index=None):
+        if allow_interrupt:
+            nodes.before_node_execution()
+        execution_block = None
+        for k, v in inputs.items():
+            if isinstance(v, ExecutionBlocker):
+                execution_block = execution_block_cb(v) if execution_block_cb else v
+                break
+        if execution_block is None:
+            if pre_execute_cb is not None and index is not None:
+                pre_execute_cb(index)
+            results.append(getattr(obj, func)(**inputs))
+        else:
+            results.append(execution_block)
+
     if input_is_list:
-        if allow_interrupt:
+        process_inputs(input_data_all, 0)
-            nodes.before_node_execution()
-        results.append(getattr(obj, func)(**input_data_all))
     elif max_len_input == 0:
-        if allow_interrupt:
+        process_inputs({})
-            nodes.before_node_execution()
+    else: 
-        results.append(getattr(obj, func)())
-    else:
         for i in range(max_len_input):
-            if allow_interrupt:
+            input_dict = slice_dict(input_data_all, i)
-                nodes.before_node_execution()
+            process_inputs(input_dict, i)
-            results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
     return results
 
-def get_output_data(obj, input_data_all):
+def merge_result_data(results, obj):
+    # check which outputs need concatenating
+    output = []
+    output_is_list = [False] * len(results[0])
+    if hasattr(obj, "OUTPUT_IS_LIST"):
+        output_is_list = obj.OUTPUT_IS_LIST
+
+    # merge node execution results
+    for i, is_list in zip(range(len(results[0])), output_is_list):
+        if is_list:
+            output.append([x for o in results for x in o[i]])
+        else:
+            output.append([o[i] for o in results])
+    return output
+
+def get_output_data(obj, input_data_all, execution_block_cb=None, pre_execute_cb=None):
     
     results = []
     uis = []
-    return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
+    subgraph_results = []
-
+    return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
-    for r in return_values:
+    has_subgraph = False
+    for i in range(len(return_values)):
+        r = return_values[i]
         if isinstance(r, dict):
             if 'ui' in r:
                 uis.append(r['ui'])
-            if 'result' in r:
+            if 'expand' in r:
-                results.append(r['result'])
+                # Perform an expansion, but do not append results
+                has_subgraph = True
+                new_graph = r['expand']
+                result = r.get("result", None)
+                if isinstance(result, ExecutionBlocker):
+                    result = tuple([result] * len(obj.RETURN_TYPES))
+                subgraph_results.append((new_graph, result))
+            elif 'result' in r:
+                result = r.get("result", None)
+                if isinstance(result, ExecutionBlocker):
+                    result = tuple([result] * len(obj.RETURN_TYPES))
+                results.append(result)
+                subgraph_results.append((None, result))
         else:
+            if isinstance(r, ExecutionBlocker):
+                r = tuple([r] * len(obj.RETURN_TYPES))
             results.append(r)
+            subgraph_results.append((None, r))
     
-    output = []
+    if has_subgraph:
-    if len(results) > 0:
+        output = subgraph_results
-        # check which outputs need concatenating
+    elif len(results) > 0:
-        output_is_list = [False] * len(results[0])
+        output = merge_result_data(results, obj)
-        if hasattr(obj, "OUTPUT_IS_LIST"):
+    else:
-            output_is_list = obj.OUTPUT_IS_LIST
+        output = []
-
-        # merge node execution results
-        for i, is_list in zip(range(len(results[0])), output_is_list):
-            if is_list:
-                output.append([x for o in results for x in o[i]])
-            else:
-                output.append([o[i] for o in results])
-
     ui = dict()    
     if len(uis) > 0:
         ui = {k: [y for x in uis for y in x[k]] for k in uis[0].keys()}
-    return output, ui
+    return output, ui, has_subgraph
 
 def format_value(x):
     if x is None:
@@ -117,53 +235,145 @@ def format_value(x):
     else:
         return str(x)
 
-def recursive_execute(server, prompt, outputs, current_item, extra_data, executed, prompt_id, outputs_ui, object_storage):
+def execute(server, dynprompt, caches, current_item, extra_data, executed, prompt_id, execution_list, pending_subgraph_results):
     unique_id = current_item
-    inputs = prompt[unique_id]['inputs']
+    real_node_id = dynprompt.get_real_node_id(unique_id)
-    class_type = prompt[unique_id]['class_type']
+    display_node_id = dynprompt.get_display_node_id(unique_id)
+    parent_node_id = dynprompt.get_parent_node_id(unique_id)
+    inputs = dynprompt.get_node(unique_id)['inputs']
+    class_type = dynprompt.get_node(unique_id)['class_type']
     class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
-    if unique_id in outputs:
+    if caches.outputs.get(unique_id) is not None:
-        return (True, None, None)
+        if server.client_id is not None:
-
+            cached_output = caches.ui.get(unique_id) or {}
-    for x in inputs:
+            server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": cached_output.get("output",None), "prompt_id": prompt_id }, server.client_id)
-        input_data = inputs[x]
+        return (ExecutionResult.SUCCESS, None, None)
-
-        if isinstance(input_data, list):
-            input_unique_id = input_data[0]
-            output_index = input_data[1]
-            if input_unique_id not in outputs:
-                result = recursive_execute(server, prompt, outputs, input_unique_id, extra_data, executed, prompt_id, outputs_ui, object_storage)
-                if result[0] is not True:
-                    # Another node failed further upstream
-                    return result
 
     input_data_all = None
     try:
-        input_data_all = get_input_data(inputs, class_def, unique_id, outputs, prompt, extra_data)
+        if unique_id in pending_subgraph_results:
-        if server.client_id is not None:
+            cached_results = pending_subgraph_results[unique_id]
-            server.last_node_id = unique_id
+            resolved_outputs = []
-            server.send_sync("executing", { "node": unique_id, "prompt_id": prompt_id }, server.client_id)
+            for is_subgraph, result in cached_results:
+                if not is_subgraph:
+                    resolved_outputs.append(result)
+                else:
+                    resolved_output = []
+                    for r in result:
+                        if is_link(r):
+                            source_node, source_output = r[0], r[1]
+                            node_output = caches.outputs.get(source_node)[source_output]
+                            for o in node_output:
+                                resolved_output.append(o)
 
-        obj = object_storage.get((unique_id, class_type), None)
+                        else:
-        if obj is None:
+                            resolved_output.append(r)
-            obj = class_def()
+                    resolved_outputs.append(tuple(resolved_output))
-            object_storage[(unique_id, class_type)] = obj
+            output_data = merge_result_data(resolved_outputs, class_def)
-
+            output_ui = []
-        output_data, output_ui = get_output_data(obj, input_data_all)
+            has_subgraph = False
-        outputs[unique_id] = output_data
+        else:
-        if len(output_ui) > 0:
+            input_data_all, missing_keys = get_input_data(inputs, class_def, unique_id, caches.outputs, dynprompt, extra_data)
-            outputs_ui[unique_id] = output_ui
             if server.client_id is not None:
-                server.send_sync("executed", { "node": unique_id, "output": output_ui, "prompt_id": prompt_id }, server.client_id)
+                server.last_node_id = display_node_id
+                server.send_sync("executing", { "node": unique_id, "display_node": display_node_id, "prompt_id": prompt_id }, server.client_id)
+
+            obj = caches.objects.get(unique_id)
+            if obj is None:
+                obj = class_def()
+                caches.objects.set(unique_id, obj)
+
+            if hasattr(obj, "check_lazy_status"):
+                required_inputs = _map_node_over_list(obj, input_data_all, "check_lazy_status", allow_interrupt=True)
+                required_inputs = set(sum([r for r in required_inputs if isinstance(r,list)], []))
+                required_inputs = [x for x in required_inputs if isinstance(x,str) and (
+                    x not in input_data_all or x in missing_keys
+                )]
+                if len(required_inputs) > 0:
+                    for i in required_inputs:
+                        execution_list.make_input_strong_link(unique_id, i)
+                    return (ExecutionResult.PENDING, None, None)
+
+            def execution_block_cb(block):
+                if block.message is not None:
+                    mes = {
+                        "prompt_id": prompt_id,
+                        "node_id": unique_id,
+                        "node_type": class_type,
+                        "executed": list(executed),
+
+                        "exception_message": f"Execution Blocked: {block.message}",
+                        "exception_type": "ExecutionBlocked",
+                        "traceback": [],
+                        "current_inputs": [],
+                        "current_outputs": [],
+                    }
+                    server.send_sync("execution_error", mes, server.client_id)
+                    return ExecutionBlocker(None)
+                else:
+                    return block
+            def pre_execute_cb(call_index):
+                GraphBuilder.set_default_prefix(unique_id, call_index, 0)
+            output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
+        if len(output_ui) > 0:
+            caches.ui.set(unique_id, {
+                "meta": {
+                    "node_id": unique_id,
+                    "display_node": display_node_id,
+                    "parent_node": parent_node_id,
+                    "real_node_id": real_node_id,
+                },
+                "output": output_ui
+            })
+            if server.client_id is not None:
+                server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": output_ui, "prompt_id": prompt_id }, server.client_id)
+        if has_subgraph:
+            cached_outputs = []
+            new_node_ids = []
+            new_output_ids = []
+            new_output_links = []
+            for i in range(len(output_data)):
+                new_graph, node_outputs = output_data[i]
+                if new_graph is None:
+                    cached_outputs.append((False, node_outputs))
+                else:
+                    # Check for conflicts
+                    for node_id in new_graph.keys():
+                        if dynprompt.has_node(node_id):
+                            raise DuplicateNodeError(f"Attempt to add duplicate node {node_id}. Ensure node ids are unique and deterministic or use graph_utils.GraphBuilder.")
+                    for node_id, node_info in new_graph.items():
+                        new_node_ids.append(node_id)
+                        display_id = node_info.get("override_display_id", unique_id)
+                        dynprompt.add_ephemeral_node(node_id, node_info, unique_id, display_id)
+                        # Figure out if the newly created node is an output node
+                        class_type = node_info["class_type"]
+                        class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
+                        if hasattr(class_def, 'OUTPUT_NODE') and class_def.OUTPUT_NODE == True:
+                            new_output_ids.append(node_id)
+                    for i in range(len(node_outputs)):
+                        if is_link(node_outputs[i]):
+                            from_node_id, from_socket = node_outputs[i][0], node_outputs[i][1]
+                            new_output_links.append((from_node_id, from_socket))
+                    cached_outputs.append((True, node_outputs))
+            new_node_ids = set(new_node_ids)
+            for cache in caches.all:
+                cache.ensure_subcache_for(unique_id, new_node_ids).clean_unused()
+            for node_id in new_output_ids:
+                execution_list.add_node(node_id)
+            for link in new_output_links:
+                execution_list.add_strong_link(link[0], link[1], unique_id)
+            pending_subgraph_results[unique_id] = cached_outputs
+            return (ExecutionResult.PENDING, None, None)
+        caches.outputs.set(unique_id, output_data)
     except comfy.model_management.InterruptProcessingException as iex:
         logging.info("Processing interrupted")
 
         # skip formatting inputs/outputs
         error_details = {
-            "node_id": unique_id,
+            "node_id": real_node_id,
         }
 
-        return (False, error_details, iex)
+        return (ExecutionResult.FAILURE, error_details, iex)
     except Exception as ex:
         typ, _, tb = sys.exc_info()
         exception_type = full_type_name(typ)
@@ -173,116 +383,36 @@ def recursive_execute(server, prompt, outputs, current_item, extra_data, execute
             for name, inputs in input_data_all.items():
                 input_data_formatted[name] = [format_value(x) for x in inputs]
 
-        output_data_formatted = {}
+        logging.error(f"!!! Exception during processing !!! {ex}")
-        for node_id, node_outputs in outputs.items():
-            output_data_formatted[node_id] = [[format_value(x) for x in l] for l in node_outputs]
-
-        logging.error(f"!!! Exception during processing!!! {ex}")
         logging.error(traceback.format_exc())
 
         error_details = {
-            "node_id": unique_id,
+            "node_id": real_node_id,
             "exception_message": str(ex),
             "exception_type": exception_type,
             "traceback": traceback.format_tb(tb),
-            "current_inputs": input_data_formatted,
+            "current_inputs": input_data_formatted
-            "current_outputs": output_data_formatted
         }
-        return (False, error_details, ex)
+        if isinstance(ex, comfy.model_management.OOM_EXCEPTION):
+            logging.error("Got an OOM, unloading all loaded models.")
+            comfy.model_management.unload_all_models()
+
+        return (ExecutionResult.FAILURE, error_details, ex)
 
     executed.add(unique_id)
 
-    return (True, None, None)
+    return (ExecutionResult.SUCCESS, None, None)
-
-def recursive_will_execute(prompt, outputs, current_item, memo={}):
-    unique_id = current_item
-
-    if unique_id in memo:
-        return memo[unique_id]
-
-    inputs = prompt[unique_id]['inputs']
-    will_execute = []
-    if unique_id in outputs:
-        return []
-
-    for x in inputs:
-        input_data = inputs[x]
-        if isinstance(input_data, list):
-            input_unique_id = input_data[0]
-            output_index = input_data[1]
-            if input_unique_id not in outputs:
-                will_execute += recursive_will_execute(prompt, outputs, input_unique_id, memo)
-
-    memo[unique_id] = will_execute + [unique_id]
-    return memo[unique_id]
-
-def recursive_output_delete_if_changed(prompt, old_prompt, outputs, current_item):
-    unique_id = current_item
-    inputs = prompt[unique_id]['inputs']
-    class_type = prompt[unique_id]['class_type']
-    class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
-
-    is_changed_old = ''
-    is_changed = ''
-    to_delete = False
-    if hasattr(class_def, 'IS_CHANGED'):
-        if unique_id in old_prompt and 'is_changed' in old_prompt[unique_id]:
-            is_changed_old = old_prompt[unique_id]['is_changed']
-        if 'is_changed' not in prompt[unique_id]:
-            input_data_all = get_input_data(inputs, class_def, unique_id, outputs)
-            if input_data_all is not None:
-                try:
-                    #is_changed = class_def.IS_CHANGED(**input_data_all)
-                    is_changed = map_node_over_list(class_def, input_data_all, "IS_CHANGED")
-                    prompt[unique_id]['is_changed'] = is_changed
-                except:
-                    to_delete = True
-        else:
-            is_changed = prompt[unique_id]['is_changed']
-
-    if unique_id not in outputs:
-        return True
-
-    if not to_delete:
-        if is_changed != is_changed_old:
-            to_delete = True
-        elif unique_id not in old_prompt:
-            to_delete = True
-        elif class_type != old_prompt[unique_id]['class_type']:
-            to_delete = True
-        elif inputs == old_prompt[unique_id]['inputs']:
-            for x in inputs:
-                input_data = inputs[x]
-
-                if isinstance(input_data, list):
-                    input_unique_id = input_data[0]
-                    output_index = input_data[1]
-                    if input_unique_id in outputs:
-                        to_delete = recursive_output_delete_if_changed(prompt, old_prompt, outputs, input_unique_id)
-                    else:
-                        to_delete = True
-                    if to_delete:
-                        break
-        else:
-            to_delete = True
-
-    if to_delete:
-        d = outputs.pop(unique_id)
-        del d
-    return to_delete
 
 class PromptExecutor:
-    def __init__(self, server):
+    def __init__(self, server, lru_size=None):
+        self.lru_size = lru_size
         self.server = server
         self.reset()
 
     def reset(self):
-        self.outputs = {}
+        self.caches = CacheSet(self.lru_size)
-        self.object_storage = {}
-        self.outputs_ui = {}
         self.status_messages = []
         self.success = True
-        self.old_prompt = {}
 
     def add_message(self, event, data: dict, broadcast: bool):
         data = {
@@ -313,26 +443,13 @@ class PromptExecutor:
                 "node_id": node_id,
                 "node_type": class_type,
                 "executed": list(executed),
-
                 "exception_message": error["exception_message"],
                 "exception_type": error["exception_type"],
                 "traceback": error["traceback"],
                 "current_inputs": error["current_inputs"],
-                "current_outputs": error["current_outputs"],
+                "current_outputs": list(current_outputs),
             }
             self.add_message("execution_error", mes, broadcast=False)
-        
-        # Next, remove the subsequent outputs since they will not be executed
-        to_delete = []
-        for o in self.outputs:
-            if (o not in current_outputs) and (o not in executed):
-                to_delete += [o]
-                if o in self.old_prompt:
-                    d = self.old_prompt.pop(o)
-                    del d
-        for o in to_delete:
-            d = self.outputs.pop(o)
-            del d
 
     def execute(self, prompt, prompt_id, extra_data={}, execute_outputs=[]):
         nodes.interrupt_processing(False)
@@ -346,65 +463,59 @@ class PromptExecutor:
         self.add_message("execution_start", { "prompt_id": prompt_id}, broadcast=False)
 
         with torch.inference_mode():
-            #delete cached outputs if nodes don't exist for them
+            dynamic_prompt = DynamicPrompt(prompt)
-            to_delete = []
+            is_changed_cache = IsChangedCache(dynamic_prompt, self.caches.outputs)
-            for o in self.outputs:
+            for cache in self.caches.all:
-                if o not in prompt:
+                cache.set_prompt(dynamic_prompt, prompt.keys(), is_changed_cache)
-                    to_delete += [o]
+                cache.clean_unused()
-            for o in to_delete:
-                d = self.outputs.pop(o)
-                del d
-            to_delete = []
-            for o in self.object_storage:
-                if o[0] not in prompt:
-                    to_delete += [o]
-                else:
-                    p = prompt[o[0]]
-                    if o[1] != p['class_type']:
-                        to_delete += [o]
-            for o in to_delete:
-                d = self.object_storage.pop(o)
-                del d
 
-            for x in prompt:
+            cached_nodes = []
-                recursive_output_delete_if_changed(prompt, self.old_prompt, self.outputs, x)
+            for node_id in prompt:
-
+                if self.caches.outputs.get(node_id) is not None:
-            current_outputs = set(self.outputs.keys())
+                    cached_nodes.append(node_id)
-            for x in list(self.outputs_ui.keys()):
-                if x not in current_outputs:
-                    d = self.outputs_ui.pop(x)
-                    del d
 
             comfy.model_management.cleanup_models(keep_clone_weights_loaded=True)
             self.add_message("execution_cached",
-                          { "nodes": list(current_outputs) , "prompt_id": prompt_id},
+                          { "nodes": cached_nodes, "prompt_id": prompt_id},
                           broadcast=False)
+            pending_subgraph_results = {}
             executed = set()
-            output_node_id = None
+            execution_list = ExecutionList(dynamic_prompt, self.caches.outputs)
-            to_execute = []
+            current_outputs = self.caches.outputs.all_node_ids()
-
             for node_id in list(execute_outputs):
-                to_execute += [(0, node_id)]
+                execution_list.add_node(node_id)
 
-            while len(to_execute) > 0:
+            while not execution_list.is_empty():
-                #always execute the output that depends on the least amount of unexecuted nodes first
+                node_id, error, ex = execution_list.stage_node_execution()
-                memo = {}
+                if error is not None:
-                to_execute = sorted(list(map(lambda a: (len(recursive_will_execute(prompt, self.outputs, a[-1], memo)), a[-1]), to_execute)))
+                    self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex)
-                output_node_id = to_execute.pop(0)[-1]
-
-                # This call shouldn't raise anything if there's an error deep in
-                # the actual SD code, instead it will report the node where the
-                # error was raised
-                self.success, error, ex = recursive_execute(self.server, prompt, self.outputs, output_node_id, extra_data, executed, prompt_id, self.outputs_ui, self.object_storage)
-                if self.success is not True:
-                    self.handle_execution_error(prompt_id, prompt, current_outputs, executed, error, ex)
                     break
+
+                result, error, ex = execute(self.server, dynamic_prompt, self.caches, node_id, extra_data, executed, prompt_id, execution_list, pending_subgraph_results)
+                self.success = result != ExecutionResult.FAILURE
+                if result == ExecutionResult.FAILURE:
+                    self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex)
+                    break
+                elif result == ExecutionResult.PENDING:
+                    execution_list.unstage_node_execution()
+                else: # result == ExecutionResult.SUCCESS:
+                    execution_list.complete_node_execution()
             else:
                 # Only execute when the while-loop ends without break
                 self.add_message("execution_success", { "prompt_id": prompt_id }, broadcast=False)
 
-            for x in executed:
+            ui_outputs = {}
-                self.old_prompt[x] = copy.deepcopy(prompt[x])
+            meta_outputs = {}
+            all_node_ids = self.caches.ui.all_node_ids()
+            for node_id in all_node_ids:
+                ui_info = self.caches.ui.get(node_id)
+                if ui_info is not None:
+                    ui_outputs[node_id] = ui_info["output"]
+                    meta_outputs[node_id] = ui_info["meta"]
+            self.history_result = {
+                "outputs": ui_outputs,
+                "meta": meta_outputs,
+            }
             self.server.last_node_id = None
             if comfy.model_management.DISABLE_SMART_MEMORY:
                 comfy.model_management.unload_all_models()
@@ -421,31 +532,37 @@ def validate_inputs(prompt, item, validated):
     obj_class = nodes.NODE_CLASS_MAPPINGS[class_type]
 
     class_inputs = obj_class.INPUT_TYPES()
-    required_inputs = class_inputs['required']
+    valid_inputs = set(class_inputs.get('required',{})).union(set(class_inputs.get('optional',{})))
 
     errors = []
     valid = True
 
     validate_function_inputs = []
+    validate_has_kwargs = False
     if hasattr(obj_class, "VALIDATE_INPUTS"):
-        validate_function_inputs = inspect.getfullargspec(obj_class.VALIDATE_INPUTS).args
+        argspec = inspect.getfullargspec(obj_class.VALIDATE_INPUTS)
+        validate_function_inputs = argspec.args
+        validate_has_kwargs = argspec.varkw is not None
+    received_types = {}
 
-    for x in required_inputs:
+    for x in valid_inputs:
+        type_input, input_category, extra_info = get_input_info(obj_class, x)
+        assert extra_info is not None
         if x not in inputs:
-            error = {
+            if input_category == "required":
-                "type": "required_input_missing",
+                error = {
-                "message": "Required input is missing",
+                    "type": "required_input_missing",
-                "details": f"{x}",
+                    "message": "Required input is missing",
-                "extra_info": {
+                    "details": f"{x}",
-                    "input_name": x
+                    "extra_info": {
+                        "input_name": x
+                    }
                 }
-            }
+                errors.append(error)
-            errors.append(error)
             continue
 
         val = inputs[x]
-        info = required_inputs[x]
+        info = (type_input, extra_info)
-        type_input = info[0]
         if isinstance(val, list):
             if len(val) != 2:
                 error = {
@@ -464,8 +581,9 @@ def validate_inputs(prompt, item, validated):
             o_id = val[0]
             o_class_type = prompt[o_id]['class_type']
             r = nodes.NODE_CLASS_MAPPINGS[o_class_type].RETURN_TYPES
-            if r[val[1]] != type_input:
+            received_type = r[val[1]]
-                received_type = r[val[1]]
+            received_types[x] = received_type
+            if 'input_types' not in validate_function_inputs and received_type != type_input:
                 details = f"{x}, {received_type} != {type_input}"
                 error = {
                     "type": "return_type_mismatch",
@@ -516,6 +634,9 @@ def validate_inputs(prompt, item, validated):
                 if type_input == "STRING":
                     val = str(val)
                     inputs[x] = val
+                if type_input == "BOOLEAN":
+                    val = bool(val)
+                    inputs[x] = val
             except Exception as ex:
                 error = {
                     "type": "invalid_input_type",
@@ -531,11 +652,11 @@ def validate_inputs(prompt, item, validated):
                 errors.append(error)
                 continue
 
-            if len(info) > 1:
+            if x not in validate_function_inputs and not validate_has_kwargs:
-                if "min" in info[1] and val < info[1]["min"]:
+                if "min" in extra_info and val < extra_info["min"]:
                     error = {
                         "type": "value_smaller_than_min",
-                        "message": "Value {} smaller than min of {}".format(val, info[1]["min"]),
+                        "message": "Value {} smaller than min of {}".format(val, extra_info["min"]),
                         "details": f"{x}",
                         "extra_info": {
                             "input_name": x,
@@ -545,10 +666,10 @@ def validate_inputs(prompt, item, validated):
                     }
                     errors.append(error)
                     continue
-                if "max" in info[1] and val > info[1]["max"]:
+                if "max" in extra_info and val > extra_info["max"]:
                     error = {
                         "type": "value_bigger_than_max",
-                        "message": "Value {} bigger than max of {}".format(val, info[1]["max"]),
+                        "message": "Value {} bigger than max of {}".format(val, extra_info["max"]),
                         "details": f"{x}",
                         "extra_info": {
                             "input_name": x,
@@ -559,7 +680,6 @@ def validate_inputs(prompt, item, validated):
                     errors.append(error)
                     continue
 
-            if x not in validate_function_inputs:
                 if isinstance(type_input, list):
                     if val not in type_input:
                         input_config = info
@@ -586,18 +706,20 @@ def validate_inputs(prompt, item, validated):
                         errors.append(error)
                         continue
 
-    if len(validate_function_inputs) > 0:
+    if len(validate_function_inputs) > 0 or validate_has_kwargs:
-        input_data_all = get_input_data(inputs, obj_class, unique_id)
+        input_data_all, _ = get_input_data(inputs, obj_class, unique_id)
         input_filtered = {}
         for x in input_data_all:
-            if x in validate_function_inputs:
+            if x in validate_function_inputs or validate_has_kwargs:
                 input_filtered[x] = input_data_all[x]
+        if 'input_types' in validate_function_inputs:
+            input_filtered['input_types'] = [received_types]
 
         #ret = obj_class.VALIDATE_INPUTS(**input_filtered)
-        ret = map_node_over_list(obj_class, input_filtered, "VALIDATE_INPUTS")
+        ret = _map_node_over_list(obj_class, input_filtered, "VALIDATE_INPUTS")
         for x in input_filtered:
             for i, r in enumerate(ret):
-                if r is not True:
+                if r is not True and not isinstance(r, ExecutionBlocker):
                     details = f"{x}"
                     if r is not False:
                         details += f" - {str(r)}"
@@ -608,8 +730,6 @@ def validate_inputs(prompt, item, validated):
                         "details": details,
                         "extra_info": {
                             "input_name": x,
-                            "input_config": info,
-                            "received_value": val,
                         }
                     }
                     errors.append(error)
@@ -775,7 +895,7 @@ class PromptQueue:
         completed: bool
         messages: List[str]
 
-    def task_done(self, item_id, outputs,
+    def task_done(self, item_id, history_result,
                   status: Optional['PromptQueue.ExecutionStatus']):
         with self.mutex:
             prompt = self.currently_running.pop(item_id)
@@ -788,9 +908,10 @@ class PromptQueue:
 
             self.history[prompt[1]] = {
                 "prompt": prompt,
-                "outputs": copy.deepcopy(outputs),
+                "outputs": {},
                 'status': status_dict,
             }
+            self.history[prompt[1]].update(history_result)
             self.server.queue_updated()
 
     def get_current_queue(self):

folder_paths.py

@@ -1,13 +1,13 @@
+from __future__ import annotations
+
 import os
 import time
 import logging
-from typing import Set, List, Dict, Tuple
+from collections.abc import Collection
 
-supported_pt_extensions: Set[str] = set(['.ckpt', '.pt', '.bin', '.pth', '.safetensors', '.pkl', '.sft'])
+supported_pt_extensions: set[str] = {'.ckpt', '.pt', '.bin', '.pth', '.safetensors', '.pkl', '.sft'}
 
-SupportedFileExtensionsType = Set[str]
+folder_names_and_paths: dict[str, tuple[list[str], set[str]]] = {}
-ScanPathType = List[str]
-folder_names_and_paths: Dict[str, Tuple[ScanPathType, SupportedFileExtensionsType]] = {}
 
 base_path = os.path.dirname(os.path.realpath(__file__))
 models_dir = os.path.join(base_path, "models")
@@ -17,7 +17,7 @@ folder_names_and_paths["configs"] = ([os.path.join(models_dir, "configs")], [".y
 folder_names_and_paths["loras"] = ([os.path.join(models_dir, "loras")], supported_pt_extensions)
 folder_names_and_paths["vae"] = ([os.path.join(models_dir, "vae")], supported_pt_extensions)
 folder_names_and_paths["clip"] = ([os.path.join(models_dir, "clip")], supported_pt_extensions)
-folder_names_and_paths["unet"] = ([os.path.join(models_dir, "unet")], supported_pt_extensions)
+folder_names_and_paths["diffusion_models"] = ([os.path.join(models_dir, "unet"), os.path.join(models_dir, "diffusion_models")], supported_pt_extensions)
 folder_names_and_paths["clip_vision"] = ([os.path.join(models_dir, "clip_vision")], supported_pt_extensions)
 folder_names_and_paths["style_models"] = ([os.path.join(models_dir, "style_models")], supported_pt_extensions)
 folder_names_and_paths["embeddings"] = ([os.path.join(models_dir, "embeddings")], supported_pt_extensions)
@@ -42,7 +42,11 @@ temp_directory = os.path.join(os.path.dirname(os.path.realpath(__file__)), "temp
 input_directory = os.path.join(os.path.dirname(os.path.realpath(__file__)), "input")
 user_directory = os.path.join(os.path.dirname(os.path.realpath(__file__)), "user")
 
-filename_list_cache = {}
+filename_list_cache: dict[str, tuple[list[str], dict[str, float], float]] = {}
+
+def map_legacy(folder_name: str) -> str:
+    legacy = {"unet": "diffusion_models"}
+    return legacy.get(folder_name, folder_name)
 
 if not os.path.exists(input_directory):
     try:
@@ -50,33 +54,33 @@ if not os.path.exists(input_directory):
     except:
         logging.error("Failed to create input directory")
 
-def set_output_directory(output_dir):
+def set_output_directory(output_dir: str) -> None:
     global output_directory
     output_directory = output_dir
 
-def set_temp_directory(temp_dir):
+def set_temp_directory(temp_dir: str) -> None:
     global temp_directory
     temp_directory = temp_dir
 
-def set_input_directory(input_dir):
+def set_input_directory(input_dir: str) -> None:
     global input_directory
     input_directory = input_dir
 
-def get_output_directory():
+def get_output_directory() -> str:
     global output_directory
     return output_directory
 
-def get_temp_directory():
+def get_temp_directory() -> str:
     global temp_directory
     return temp_directory
 
-def get_input_directory():
+def get_input_directory() -> str:
     global input_directory
     return input_directory
 
 
 #NOTE: used in http server so don't put folders that should not be accessed remotely
-def get_directory_by_type(type_name):
+def get_directory_by_type(type_name: str) -> str | None:
     if type_name == "output":
         return get_output_directory()
     if type_name == "temp":
@@ -88,7 +92,7 @@ def get_directory_by_type(type_name):
 
 # determine base_dir rely on annotation if name is 'filename.ext [annotation]' format
 # otherwise use default_path as base_dir
-def annotated_filepath(name):
+def annotated_filepath(name: str) -> tuple[str, str | None]:
     if name.endswith("[output]"):
         base_dir = get_output_directory()
         name = name[:-9]
@@ -104,7 +108,7 @@ def annotated_filepath(name):
     return name, base_dir
 
 
-def get_annotated_filepath(name, default_dir=None):
+def get_annotated_filepath(name: str, default_dir: str | None=None) -> str:
     name, base_dir = annotated_filepath(name)
 
     if base_dir is None:
@@ -116,7 +120,7 @@ def get_annotated_filepath(name, default_dir=None):
     return os.path.join(base_dir, name)
 
 
-def exists_annotated_filepath(name):
+def exists_annotated_filepath(name) -> bool:
     name, base_dir = annotated_filepath(name)
 
     if base_dir is None:
@@ -126,17 +130,19 @@ def exists_annotated_filepath(name):
     return os.path.exists(filepath)
 
 
-def add_model_folder_path(folder_name, full_folder_path):
+def add_model_folder_path(folder_name: str, full_folder_path: str) -> None:
     global folder_names_and_paths
+    folder_name = map_legacy(folder_name)
     if folder_name in folder_names_and_paths:
         folder_names_and_paths[folder_name][0].append(full_folder_path)
     else:
         folder_names_and_paths[folder_name] = ([full_folder_path], set())
 
-def get_folder_paths(folder_name):
+def get_folder_paths(folder_name: str) -> list[str]:
+    folder_name = map_legacy(folder_name)
     return folder_names_and_paths[folder_name][0][:]
 
-def recursive_search(directory, excluded_dir_names=None):
+def recursive_search(directory: str, excluded_dir_names: list[str] | None=None) -> tuple[list[str], dict[str, float]]:
     if not os.path.isdir(directory):
         return [], {}
 
@@ -153,6 +159,10 @@ def recursive_search(directory, excluded_dir_names=None):
         logging.warning(f"Warning: Unable to access {directory}. Skipping this path.")
 
     logging.debug("recursive file list on directory {}".format(directory))
+    dirpath: str
+    subdirs: list[str]
+    filenames: list[str]
+
     for dirpath, subdirs, filenames in os.walk(directory, followlinks=True, topdown=True):
         subdirs[:] = [d for d in subdirs if d not in excluded_dir_names]
         for file_name in filenames:
@@ -160,7 +170,7 @@ def recursive_search(directory, excluded_dir_names=None):
             result.append(relative_path)
 
         for d in subdirs:
-            path = os.path.join(dirpath, d)
+            path: str = os.path.join(dirpath, d)
             try:
                 dirs[path] = os.path.getmtime(path)
             except FileNotFoundError:
@@ -169,13 +179,14 @@ def recursive_search(directory, excluded_dir_names=None):
     logging.debug("found {} files".format(len(result)))
     return result, dirs
 
-def filter_files_extensions(files, extensions):
+def filter_files_extensions(files: Collection[str], extensions: Collection[str]) -> list[str]:
     return sorted(list(filter(lambda a: os.path.splitext(a)[-1].lower() in extensions or len(extensions) == 0, files)))
 
 
 
-def get_full_path(folder_name, filename):
+def get_full_path(folder_name: str, filename: str) -> str | None:
     global folder_names_and_paths
+    folder_name = map_legacy(folder_name)
     if folder_name not in folder_names_and_paths:
         return None
     folders = folder_names_and_paths[folder_name]
@@ -189,7 +200,8 @@ def get_full_path(folder_name, filename):
 
     return None
 
-def get_filename_list_(folder_name):
+def get_filename_list_(folder_name: str) -> tuple[list[str], dict[str, float], float]:
+    folder_name = map_legacy(folder_name)
     global folder_names_and_paths
     output_list = set()
     folders = folder_names_and_paths[folder_name]
@@ -199,11 +211,12 @@ def get_filename_list_(folder_name):
         output_list.update(filter_files_extensions(files, folders[1]))
         output_folders = {**output_folders, **folders_all}
 
-    return (sorted(list(output_list)), output_folders, time.perf_counter())
+    return sorted(list(output_list)), output_folders, time.perf_counter()
 
-def cached_filename_list_(folder_name):
+def cached_filename_list_(folder_name: str) -> tuple[list[str], dict[str, float], float] | None:
     global filename_list_cache
     global folder_names_and_paths
+    folder_name = map_legacy(folder_name)
     if folder_name not in filename_list_cache:
         return None
     out = filename_list_cache[folder_name]
@@ -222,7 +235,8 @@ def cached_filename_list_(folder_name):
 
     return out
 
-def get_filename_list(folder_name):
+def get_filename_list(folder_name: str) -> list[str]:
+    folder_name = map_legacy(folder_name)
     out = cached_filename_list_(folder_name)
     if out is None:
         out = get_filename_list_(folder_name)
@@ -230,17 +244,17 @@ def get_filename_list(folder_name):
         filename_list_cache[folder_name] = out
     return list(out[0])
 
-def get_save_image_path(filename_prefix, output_dir, image_width=0, image_height=0):
+def get_save_image_path(filename_prefix: str, output_dir: str, image_width=0, image_height=0) -> tuple[str, str, int, str, str]:
-    def map_filename(filename):
+    def map_filename(filename: str) -> tuple[int, str]:
         prefix_len = len(os.path.basename(filename_prefix))
         prefix = filename[:prefix_len + 1]
         try:
             digits = int(filename[prefix_len + 1:].split('_')[0])
         except:
             digits = 0
-        return (digits, prefix)
+        return digits, prefix
 
-    def compute_vars(input, image_width, image_height):
+    def compute_vars(input: str, image_width: int, image_height: int) -> str:
         input = input.replace("%width%", str(image_width))
         input = input.replace("%height%", str(image_height))
         return input

main.py

@@ -101,7 +101,7 @@ def cuda_malloc_warning():
             logging.warning("\nWARNING: this card most likely does not support cuda-malloc, if you get \"CUDA error\" please run ComfyUI with: --disable-cuda-malloc\n")
 
 def prompt_worker(q, server):
-    e = execution.PromptExecutor(server)
+    e = execution.PromptExecutor(server, lru_size=args.cache_lru)
     last_gc_collect = 0
     need_gc = False
     gc_collect_interval = 10.0
@@ -121,7 +121,7 @@ def prompt_worker(q, server):
             e.execute(item[2], prompt_id, item[3], item[4])
             need_gc = True
             q.task_done(item_id,
-                        e.outputs_ui,
+                        e.history_result,
                         status=execution.PromptQueue.ExecutionStatus(
                             status_str='success' if e.success else 'error',
                             completed=e.success,
@@ -242,6 +242,7 @@ if __name__ == "__main__":
     folder_paths.add_model_folder_path("checkpoints", os.path.join(folder_paths.get_output_directory(), "checkpoints"))
     folder_paths.add_model_folder_path("clip", os.path.join(folder_paths.get_output_directory(), "clip"))
     folder_paths.add_model_folder_path("vae", os.path.join(folder_paths.get_output_directory(), "vae"))
+    folder_paths.add_model_folder_path("diffusion_models", os.path.join(folder_paths.get_output_directory(), "diffusion_models"))
 
     if args.input_directory:
         input_dir = os.path.abspath(args.input_directory)
@@ -261,6 +262,7 @@ if __name__ == "__main__":
         call_on_start = startup_server
 
     try:
+        loop.run_until_complete(server.setup())
         loop.run_until_complete(run(server, address=args.listen, port=args.port, verbose=not args.dont_print_server, call_on_start=call_on_start))
     except KeyboardInterrupt:
         logging.info("\nStopped server")

model_filemanager/__init__.py

@@ -0,0 +1,2 @@
+# model_manager/__init__.py
+from .download_models import download_model, DownloadModelStatus, DownloadStatusType, create_model_path, check_file_exists, track_download_progress, validate_model_subdirectory, validate_filename

model_filemanager/download_models.py

@@ -0,0 +1,240 @@
+from __future__ import annotations
+import aiohttp
+import os
+import traceback
+import logging
+from folder_paths import models_dir
+import re
+from typing import Callable, Any, Optional, Awaitable, Dict
+from enum import Enum
+import time
+from dataclasses import dataclass
+
+
+class DownloadStatusType(Enum):
+    PENDING = "pending"
+    IN_PROGRESS = "in_progress"
+    COMPLETED = "completed"
+    ERROR = "error"
+
+@dataclass
+class DownloadModelStatus():
+    status: str
+    progress_percentage: float
+    message: str
+    already_existed: bool = False
+
+    def __init__(self, status: DownloadStatusType, progress_percentage: float, message: str, already_existed: bool):
+        self.status = status.value  # Store the string value of the Enum
+        self.progress_percentage = progress_percentage
+        self.message = message
+        self.already_existed = already_existed
+    
+    def to_dict(self) -> Dict[str, Any]:
+        return {
+            "status": self.status,
+            "progress_percentage": self.progress_percentage,
+            "message": self.message,
+            "already_existed": self.already_existed
+        }
+
+async def download_model(model_download_request: Callable[[str], Awaitable[aiohttp.ClientResponse]],
+                         model_name: str,  
+                         model_url: str, 
+                         model_sub_directory: str,
+                         progress_callback: Callable[[str, DownloadModelStatus], Awaitable[Any]],
+                         progress_interval: float = 1.0) -> DownloadModelStatus:
+    """
+    Download a model file from a given URL into the models directory.
+
+    Args:
+        model_download_request (Callable[[str], Awaitable[aiohttp.ClientResponse]]): 
+            A function that makes an HTTP request. This makes it easier to mock in unit tests.
+        model_name (str): 
+            The name of the model file to be downloaded. This will be the filename on disk.
+        model_url (str): 
+            The URL from which to download the model.
+        model_sub_directory (str): 
+            The subdirectory within the main models directory where the model 
+            should be saved (e.g., 'checkpoints', 'loras', etc.).
+        progress_callback (Callable[[str, DownloadModelStatus], Awaitable[Any]]): 
+            An asynchronous function to call with progress updates.
+
+    Returns:
+        DownloadModelStatus: The result of the download operation.
+    """
+    if not validate_model_subdirectory(model_sub_directory):
+        return DownloadModelStatus(
+            DownloadStatusType.ERROR, 
+            0,
+            "Invalid model subdirectory", 
+            False
+        )
+
+    if not validate_filename(model_name):
+        return DownloadModelStatus(
+            DownloadStatusType.ERROR, 
+            0,
+            "Invalid model name", 
+            False
+        )
+
+    file_path, relative_path = create_model_path(model_name, model_sub_directory, models_dir)
+    existing_file = await check_file_exists(file_path, model_name, progress_callback, relative_path)
+    if existing_file:
+        return existing_file
+
+    try:
+        status = DownloadModelStatus(DownloadStatusType.PENDING, 0, f"Starting download of {model_name}", False)
+        await progress_callback(relative_path, status)
+
+        response = await model_download_request(model_url)
+        if response.status != 200:
+            error_message = f"Failed to download {model_name}. Status code: {response.status}"
+            logging.error(error_message)
+            status = DownloadModelStatus(DownloadStatusType.ERROR, 0, error_message, False)
+            await progress_callback(relative_path, status)
+            return DownloadModelStatus(DownloadStatusType.ERROR, 0, error_message, False)
+
+        return await track_download_progress(response, file_path, model_name, progress_callback, relative_path, progress_interval)
+
+    except Exception as e:
+        logging.error(f"Error in downloading model: {e}")
+        return await handle_download_error(e, model_name, progress_callback, relative_path)
+    
+
+def create_model_path(model_name: str, model_directory: str, models_base_dir: str) -> tuple[str, str]:
+    full_model_dir = os.path.join(models_base_dir, model_directory)
+    os.makedirs(full_model_dir, exist_ok=True)
+    file_path = os.path.join(full_model_dir, model_name)
+
+    # Ensure the resulting path is still within the base directory
+    abs_file_path = os.path.abspath(file_path)
+    abs_base_dir = os.path.abspath(str(models_base_dir))
+    if os.path.commonprefix([abs_file_path, abs_base_dir]) != abs_base_dir:
+        raise Exception(f"Invalid model directory: {model_directory}/{model_name}")
+
+
+    relative_path = '/'.join([model_directory, model_name])
+    return file_path, relative_path
+
+async def check_file_exists(file_path: str, 
+                            model_name: str, 
+                            progress_callback: Callable[[str, DownloadModelStatus], Awaitable[Any]], 
+                            relative_path: str) -> Optional[DownloadModelStatus]:
+    if os.path.exists(file_path):
+        status = DownloadModelStatus(DownloadStatusType.COMPLETED, 100, f"{model_name} already exists", True)
+        await progress_callback(relative_path, status)
+        return status
+    return None
+
+
+async def track_download_progress(response: aiohttp.ClientResponse, 
+                                  file_path: str, 
+                                  model_name: str, 
+                                  progress_callback: Callable[[str, DownloadModelStatus], Awaitable[Any]], 
+                                  relative_path: str, 
+                                  interval: float = 1.0) -> DownloadModelStatus:
+    try:
+        total_size = int(response.headers.get('Content-Length', 0))
+        downloaded = 0
+        last_update_time = time.time()
+
+        async def update_progress():
+            nonlocal last_update_time
+            progress = (downloaded / total_size) * 100 if total_size > 0 else 0
+            status = DownloadModelStatus(DownloadStatusType.IN_PROGRESS, progress, f"Downloading {model_name}", False)
+            await progress_callback(relative_path, status)
+            last_update_time = time.time()
+
+        with open(file_path, 'wb') as f:
+            chunk_iterator = response.content.iter_chunked(8192)
+            while True:
+                try:
+                    chunk = await chunk_iterator.__anext__()
+                except StopAsyncIteration:
+                    break
+                f.write(chunk)
+                downloaded += len(chunk)
+                
+                if time.time() - last_update_time >= interval:
+                    await update_progress()
+
+        await update_progress()
+        
+        logging.info(f"Successfully downloaded {model_name}. Total downloaded: {downloaded}")
+        status = DownloadModelStatus(DownloadStatusType.COMPLETED, 100, f"Successfully downloaded {model_name}", False)
+        await progress_callback(relative_path, status)
+
+        return status
+    except Exception as e:
+        logging.error(f"Error in track_download_progress: {e}")
+        logging.error(traceback.format_exc())
+        return await handle_download_error(e, model_name, progress_callback, relative_path)
+
+async def handle_download_error(e: Exception, 
+                                model_name: str, 
+                                progress_callback: Callable[[str, DownloadModelStatus], Any], 
+                                relative_path: str) -> DownloadModelStatus:
+    error_message = f"Error downloading {model_name}: {str(e)}"
+    status = DownloadModelStatus(DownloadStatusType.ERROR, 0, error_message, False)
+    await progress_callback(relative_path, status)
+    return status
+
+def validate_model_subdirectory(model_subdirectory: str) -> bool:
+    """
+    Validate that the model subdirectory is safe to install into. 
+    Must not contain relative paths, nested paths or special characters
+    other than underscores and hyphens.
+
+    Args:
+        model_subdirectory (str): The subdirectory for the specific model type.
+
+    Returns:
+        bool: True if the subdirectory is safe, False otherwise.
+    """
+    if len(model_subdirectory) > 50:
+        return False
+
+    if '..' in model_subdirectory or '/' in model_subdirectory:
+        return False
+
+    if not re.match(r'^[a-zA-Z0-9_-]+$', model_subdirectory):
+        return False
+
+    return True
+
+def validate_filename(filename: str)-> bool:
+    """
+    Validate a filename to ensure it's safe and doesn't contain any path traversal attempts.
+    
+    Args:
+    filename (str): The filename to validate
+
+    Returns:
+    bool: True if the filename is valid, False otherwise
+    """
+    if not filename.lower().endswith(('.sft', '.safetensors')):
+        return False
+
+    # Check if the filename is empty, None, or just whitespace
+    if not filename or not filename.strip():
+        return False
+
+    # Check for any directory traversal attempts or invalid characters
+    if any(char in filename for char in ['..', '/', '\\', '\n', '\r', '\t', '\0']):
+        return False
+
+    # Check if the filename starts with a dot (hidden file)
+    if filename.startswith('.'):
+        return False
+
+    # Use a whitelist of allowed characters
+    if not re.match(r'^[a-zA-Z0-9_\-. ]+$', filename):
+        return False
+
+    # Ensure the filename isn't too long
+    if len(filename) > 255:
+        return False
+
+    return True

nodes.py

@@ -47,11 +47,18 @@ MAX_RESOLUTION=16384
 class CLIPTextEncode:
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": {"text": ("STRING", {"multiline": True, "dynamicPrompts": True}), "clip": ("CLIP", )}}
+        return {
+            "required": {
+                "text": ("STRING", {"multiline": True, "dynamicPrompts": True, "tooltip": "The text to be encoded."}), 
+                "clip": ("CLIP", {"tooltip": "The CLIP model used for encoding the text."})
+            }
+        }
     RETURN_TYPES = ("CONDITIONING",)
+    OUTPUT_TOOLTIPS = ("A conditioning containing the embedded text used to guide the diffusion model.",)
     FUNCTION = "encode"
 
     CATEGORY = "conditioning"
+    DESCRIPTION = "Encodes a text prompt using a CLIP model into an embedding that can be used to guide the diffusion model towards generating specific images."
 
     def encode(self, clip, text):
         tokens = clip.tokenize(text)
@@ -260,11 +267,18 @@ class ConditioningSetTimestepRange:
 class VAEDecode:
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": { "samples": ("LATENT", ), "vae": ("VAE", )}}
+        return {
+            "required": { 
+                "samples": ("LATENT", {"tooltip": "The latent to be decoded."}), 
+                "vae": ("VAE", {"tooltip": "The VAE model used for decoding the latent."})
+            }
+        }
     RETURN_TYPES = ("IMAGE",)
+    OUTPUT_TOOLTIPS = ("The decoded image.",)
     FUNCTION = "decode"
 
     CATEGORY = "latent"
+    DESCRIPTION = "Decodes latent images back into pixel space images."
 
     def decode(self, vae, samples):
         return (vae.decode(samples["samples"]), )
@@ -506,12 +520,19 @@ class CheckpointLoader:
 class CheckpointLoaderSimple:
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": { "ckpt_name": (folder_paths.get_filename_list("checkpoints"), ),
+        return {
-                             }}
+            "required": { 
+                "ckpt_name": (folder_paths.get_filename_list("checkpoints"), {"tooltip": "The name of the checkpoint (model) to load."}),
+            }
+        }
     RETURN_TYPES = ("MODEL", "CLIP", "VAE")
+    OUTPUT_TOOLTIPS = ("The model used for denoising latents.", 
+                       "The CLIP model used for encoding text prompts.", 
+                       "The VAE model used for encoding and decoding images to and from latent space.")
     FUNCTION = "load_checkpoint"
 
     CATEGORY = "loaders"
+    DESCRIPTION = "Loads a diffusion model checkpoint, diffusion models are used to denoise latents."
 
     def load_checkpoint(self, ckpt_name):
         ckpt_path = folder_paths.get_full_path("checkpoints", ckpt_name)
@@ -582,16 +603,22 @@ class LoraLoader:
 
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": { "model": ("MODEL",),
+        return {
-                              "clip": ("CLIP", ),
+            "required": { 
-                              "lora_name": (folder_paths.get_filename_list("loras"), ),
+                "model": ("MODEL", {"tooltip": "The diffusion model the LoRA will be applied to."}),
-                              "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01}),
+                "clip": ("CLIP", {"tooltip": "The CLIP model the LoRA will be applied to."}),
-                              "strength_clip": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01}),
+                "lora_name": (folder_paths.get_filename_list("loras"), {"tooltip": "The name of the LoRA."}),
-                              }}
+                "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01, "tooltip": "How strongly to modify the diffusion model. This value can be negative."}),
+                "strength_clip": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01, "tooltip": "How strongly to modify the CLIP model. This value can be negative."}),
+            }
+        }
+    
     RETURN_TYPES = ("MODEL", "CLIP")
+    OUTPUT_TOOLTIPS = ("The modified diffusion model.", "The modified CLIP model.")
     FUNCTION = "load_lora"
 
     CATEGORY = "loaders"
+    DESCRIPTION = "LoRAs are used to modify diffusion and CLIP models, altering the way in which latents are denoised such as applying styles. Multiple LoRA nodes can be linked together."
 
     def load_lora(self, model, clip, lora_name, strength_model, strength_clip):
         if strength_model == 0 and strength_clip == 0:
@@ -638,6 +665,8 @@ class VAELoader:
         sd1_taesd_dec = False
         sd3_taesd_enc = False
         sd3_taesd_dec = False
+        f1_taesd_enc = False
+        f1_taesd_dec = False
 
         for v in approx_vaes:
             if v.startswith("taesd_decoder."):
@@ -652,12 +681,18 @@ class VAELoader:
                 sd3_taesd_dec = True
             elif v.startswith("taesd3_encoder."):
                 sd3_taesd_enc = True
+            elif v.startswith("taef1_encoder."):
+                f1_taesd_dec = True
+            elif v.startswith("taef1_decoder."):
+                f1_taesd_enc = True
         if sd1_taesd_dec and sd1_taesd_enc:
             vaes.append("taesd")
         if sdxl_taesd_dec and sdxl_taesd_enc:
             vaes.append("taesdxl")
         if sd3_taesd_dec and sd3_taesd_enc:
             vaes.append("taesd3")
+        if f1_taesd_dec and f1_taesd_enc:
+            vaes.append("taef1")
         return vaes
 
     @staticmethod
@@ -685,6 +720,9 @@ class VAELoader:
         elif name == "taesd3":
             sd["vae_scale"] = torch.tensor(1.5305)
             sd["vae_shift"] = torch.tensor(0.0609)
+        elif name == "taef1":
+            sd["vae_scale"] = torch.tensor(0.3611)
+            sd["vae_shift"] = torch.tensor(0.1159)
         return sd
 
     @classmethod
@@ -697,7 +735,7 @@ class VAELoader:
 
     #TODO: scale factor?
     def load_vae(self, vae_name):
-        if vae_name in ["taesd", "taesdxl", "taesd3"]:
+        if vae_name in ["taesd", "taesdxl", "taesd3", "taef1"]:
             sd = self.load_taesd(vae_name)
         else:
             vae_path = folder_paths.get_full_path("vae", vae_name)
@@ -817,7 +855,7 @@ class ControlNetApplyAdvanced:
 class UNETLoader:
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": { "unet_name": (folder_paths.get_filename_list("unet"), ),
+        return {"required": { "unet_name": (folder_paths.get_filename_list("diffusion_models"), ),
                               "weight_dtype": (["default", "fp8_e4m3fn", "fp8_e5m2"],)
                              }}
     RETURN_TYPES = ("MODEL",)
@@ -826,14 +864,14 @@ class UNETLoader:
     CATEGORY = "advanced/loaders"
 
     def load_unet(self, unet_name, weight_dtype):
-        dtype = None
+        model_options = {}
         if weight_dtype == "fp8_e4m3fn":
-            dtype = torch.float8_e4m3fn
+            model_options["dtype"] = torch.float8_e4m3fn
         elif weight_dtype == "fp8_e5m2":
-            dtype = torch.float8_e5m2
+            model_options["dtype"] = torch.float8_e5m2
 
-        unet_path = folder_paths.get_full_path("unet", unet_name)
+        unet_path = folder_paths.get_full_path("diffusion_models", unet_name)
-        model = comfy.sd.load_unet(unet_path, dtype=dtype)
+        model = comfy.sd.load_diffusion_model(unet_path, model_options=model_options)
         return (model,)
 
 class CLIPLoader:
@@ -1033,13 +1071,19 @@ class EmptyLatentImage:
 
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": { "width": ("INT", {"default": 512, "min": 16, "max": MAX_RESOLUTION, "step": 8}),
+        return {
-                              "height": ("INT", {"default": 512, "min": 16, "max": MAX_RESOLUTION, "step": 8}),
+            "required": { 
-                              "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096})}}
+                "width": ("INT", {"default": 512, "min": 16, "max": MAX_RESOLUTION, "step": 8, "tooltip": "The width of the latent images in pixels."}),
+                "height": ("INT", {"default": 512, "min": 16, "max": MAX_RESOLUTION, "step": 8, "tooltip": "The height of the latent images in pixels."}),
+                "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096, "tooltip": "The number of latent images in the batch."})
+            }
+        }
     RETURN_TYPES = ("LATENT",)
+    OUTPUT_TOOLTIPS = ("The empty latent image batch.",)
     FUNCTION = "generate"
 
     CATEGORY = "latent"
+    DESCRIPTION = "Create a new batch of empty latent images to be denoised via sampling."
 
     def generate(self, width, height, batch_size=1):
         latent = torch.zeros([batch_size, 4, height // 8, width // 8], device=self.device)
@@ -1359,24 +1403,27 @@ def common_ksampler(model, seed, steps, cfg, sampler_name, scheduler, positive,
 class KSampler:
     @classmethod
     def INPUT_TYPES(s):
-        return {"required":
+        return {
-                    {"model": ("MODEL",),
+            "required": {
-                    "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
+                "model": ("MODEL", {"tooltip": "The model used for denoising the input latent."}),
-                    "steps": ("INT", {"default": 20, "min": 1, "max": 10000}),
+                "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff, "tooltip": "The random seed used for creating the noise."}),
-                    "cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}),
+                "steps": ("INT", {"default": 20, "min": 1, "max": 10000, "tooltip": "The number of steps used in the denoising process."}),
-                    "sampler_name": (comfy.samplers.KSampler.SAMPLERS, ),
+                "cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01, "tooltip": "The Classifier-Free Guidance scale balances creativity and adherence to the prompt. Higher values result in images more closely matching the prompt however too high values will negatively impact quality."}),
-                    "scheduler": (comfy.samplers.KSampler.SCHEDULERS, ),
+                "sampler_name": (comfy.samplers.KSampler.SAMPLERS, {"tooltip": "The algorithm used when sampling, this can affect the quality, speed, and style of the generated output."}),
-                    "positive": ("CONDITIONING", ),
+                "scheduler": (comfy.samplers.KSampler.SCHEDULERS, {"tooltip": "The scheduler controls how noise is gradually removed to form the image."}),
-                    "negative": ("CONDITIONING", ),
+                "positive": ("CONDITIONING", {"tooltip": "The conditioning describing the attributes you want to include in the image."}),
-                    "latent_image": ("LATENT", ),
+                "negative": ("CONDITIONING", {"tooltip": "The conditioning describing the attributes you want to exclude from the image."}),
-                    "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "latent_image": ("LATENT", {"tooltip": "The latent image to denoise."}),
-                     }
+                "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01, "tooltip": "The amount of denoising applied, lower values will maintain the structure of the initial image allowing for image to image sampling."}),
-                }
+            }
+        }
 
     RETURN_TYPES = ("LATENT",)
+    OUTPUT_TOOLTIPS = ("The denoised latent.",)
     FUNCTION = "sample"
 
     CATEGORY = "sampling"
+    DESCRIPTION = "Uses the provided model, positive and negative conditioning to denoise the latent image."
 
     def sample(self, model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=1.0):
         return common_ksampler(model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=denoise)
@@ -1424,11 +1471,15 @@ class SaveImage:
 
     @classmethod
     def INPUT_TYPES(s):
-        return {"required": 
+        return {
-                    {"images": ("IMAGE", ),
+            "required": {
-                     "filename_prefix": ("STRING", {"default": "ComfyUI"})},
+                "images": ("IMAGE", {"tooltip": "The images to save."}),
-                "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"},
+                "filename_prefix": ("STRING", {"default": "ComfyUI", "tooltip": "The prefix for the file to save. This may include formatting information such as %date:yyyy-MM-dd% or %Empty Latent Image.width% to include values from nodes."})
-                }
+            },
+            "hidden": {
+                "prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"
+            },
+        }
 
     RETURN_TYPES = ()
     FUNCTION = "save_images"
@@ -1436,6 +1487,7 @@ class SaveImage:
     OUTPUT_NODE = True
 
     CATEGORY = "image"
+    DESCRIPTION = "Saves the input images to your ComfyUI output directory."
 
     def save_images(self, images, filename_prefix="ComfyUI", prompt=None, extra_pnginfo=None):
         filename_prefix += self.prefix_append

pytest.ini

@@ -1,6 +1,7 @@
 [pytest]
 markers = 
   inference: mark as inference test (deselect with '-m "not inference"')
+  execution: mark as execution test (deselect with '-m "not execution"')
 testpaths =
   tests
   tests-unit

server.py

@@ -12,7 +12,6 @@ import json
 import glob
 import struct
 import ssl
-import hashlib
 from PIL import Image, ImageOps
 from PIL.PngImagePlugin import PngInfo
 from io import BytesIO
@@ -28,6 +27,9 @@ import comfy.model_management
 import node_helpers
 from app.frontend_management import FrontendManager
 from app.user_manager import UserManager
+from model_filemanager import download_model, DownloadModelStatus
+from typing import Optional
+from api_server.routes.internal.internal_routes import InternalRoutes
 
 
 class BinaryEventTypes:
@@ -72,10 +74,12 @@ class PromptServer():
         mimetypes.types_map['.js'] = 'application/javascript; charset=utf-8'
 
         self.user_manager = UserManager()
+        self.internal_routes = InternalRoutes()
         self.supports = ["custom_nodes_from_web"]
         self.prompt_queue = None
         self.loop = loop
         self.messages = asyncio.Queue()
+        self.client_session:Optional[aiohttp.ClientSession] = None
         self.number = 0
 
         middlewares = [cache_control]
@@ -127,13 +131,25 @@ class PromptServer():
 
         @routes.get("/")
         async def get_root(request):
-            return web.FileResponse(os.path.join(self.web_root, "index.html"))
+            response = web.FileResponse(os.path.join(self.web_root, "index.html"))
+            response.headers['Cache-Control'] = 'no-cache'
+            response.headers["Pragma"] = "no-cache"
+            response.headers["Expires"] = "0"
+            return response
 
         @routes.get("/embeddings")
         def get_embeddings(self):
             embeddings = folder_paths.get_filename_list("embeddings")
             return web.json_response(list(map(lambda a: os.path.splitext(a)[0], embeddings)))
 
+        @routes.get("/models/{folder}")
+        async def get_models(request):
+            folder = request.match_info.get("folder", None)
+            if not folder in folder_paths.folder_names_and_paths:
+                return web.Response(status=404)
+            files = folder_paths.get_filename_list(folder)
+            return web.json_response(files)
+
         @routes.get("/extensions")
         async def get_extensions(request):
             files = glob.glob(os.path.join(
@@ -418,6 +434,7 @@ class PromptServer():
             obj_class = nodes.NODE_CLASS_MAPPINGS[node_class]
             info = {}
             info['input'] = obj_class.INPUT_TYPES()
+            info['input_order'] = {key: list(value.keys()) for (key, value) in obj_class.INPUT_TYPES().items()}
             info['output'] = obj_class.RETURN_TYPES
             info['output_is_list'] = obj_class.OUTPUT_IS_LIST if hasattr(obj_class, 'OUTPUT_IS_LIST') else [False] * len(obj_class.RETURN_TYPES)
             info['output_name'] = obj_class.RETURN_NAMES if hasattr(obj_class, 'RETURN_NAMES') else info['output']
@@ -433,6 +450,14 @@ class PromptServer():
 
             if hasattr(obj_class, 'CATEGORY'):
                 info['category'] = obj_class.CATEGORY
+
+            if hasattr(obj_class, 'OUTPUT_TOOLTIPS'):
+                info['output_tooltips'] = obj_class.OUTPUT_TOOLTIPS
+
+            if getattr(obj_class, "DEPRECATED", False):
+                info['deprecated'] = True
+            if getattr(obj_class, "EXPERIMENTAL", False):
+                info['experimental'] = True
             return info
 
         @routes.get("/object_info")
@@ -555,9 +580,42 @@ class PromptServer():
                     self.prompt_queue.delete_history_item(id_to_delete)
 
             return web.Response(status=200)
+        
+        # Internal route. Should not be depended upon and is subject to change at any time.
+        # TODO(robinhuang): Move to internal route table class once we refactor PromptServer to pass around Websocket.
+        @routes.post("/internal/models/download")
+        async def download_handler(request):
+            async def report_progress(filename: str, status: DownloadModelStatus):
+                payload = status.to_dict()
+                payload['download_path'] = filename
+                await self.send_json("download_progress", payload)
+
+            data = await request.json()
+            url = data.get('url')
+            model_directory = data.get('model_directory')
+            model_filename = data.get('model_filename')
+            progress_interval = data.get('progress_interval', 1.0) # In seconds, how often to report download progress.
+
+            if not url or not model_directory or not model_filename:
+                return web.json_response({"status": "error", "message": "Missing URL or folder path or filename"}, status=400)
+
+            session = self.client_session
+            if session is None:
+                logging.error("Client session is not initialized")
+                return web.Response(status=500)
+            
+            task = asyncio.create_task(download_model(lambda url: session.get(url), model_filename, url, model_directory, report_progress, progress_interval))
+            await task
+
+            return web.json_response(task.result().to_dict())
+
+    async def setup(self):
+        timeout = aiohttp.ClientTimeout(total=None) # no timeout
+        self.client_session = aiohttp.ClientSession(timeout=timeout)
 
     def add_routes(self):
         self.user_manager.add_routes(self.routes)
+        self.app.add_subapp('/internal', self.internal_routes.get_app())
 
         # Prefix every route with /api for easier matching for delegation.
         # This is very useful for frontend dev server, which need to forward
@@ -676,6 +734,9 @@ class PromptServer():
         site = web.TCPSite(runner, address, port, ssl_context=ssl_ctx)
         await site.start()
 
+        self.address = address
+        self.port = port
+
         if verbose:
             logging.info("Starting server\n")
             logging.info("To see the GUI go to: {}://{}:{}".format(scheme, address, port))

tests-unit/app_test/frontend_manager_test.py

@@ -1,6 +1,7 @@
 import argparse
 import pytest
 from requests.exceptions import HTTPError
+from unittest.mock import patch
 
 from app.frontend_management import (
     FrontendManager,
@@ -83,6 +84,35 @@ def test_init_frontend_invalid_provider():
     with pytest.raises(HTTPError):
         FrontendManager.init_frontend_unsafe(version_string)
 
+@pytest.fixture
+def mock_os_functions():
+    with patch('app.frontend_management.os.makedirs') as mock_makedirs, \
+         patch('app.frontend_management.os.listdir') as mock_listdir, \
+         patch('app.frontend_management.os.rmdir') as mock_rmdir:
+        mock_listdir.return_value = []  # Simulate empty directory
+        yield mock_makedirs, mock_listdir, mock_rmdir
+
+@pytest.fixture
+def mock_download():
+    with patch('app.frontend_management.download_release_asset_zip') as mock:
+        mock.side_effect = Exception("Download failed")  # Simulate download failure
+        yield mock
+
+def test_finally_block(mock_os_functions, mock_download, mock_provider):
+    # Arrange
+    mock_makedirs, mock_listdir, mock_rmdir = mock_os_functions
+    version_string = 'test-owner/test-repo@1.0.0'
+
+    # Act & Assert
+    with pytest.raises(Exception):
+        FrontendManager.init_frontend_unsafe(version_string, mock_provider)
+
+    # Assert
+    mock_makedirs.assert_called_once()
+    mock_download.assert_called_once()
+    mock_listdir.assert_called_once()
+    mock_rmdir.assert_called_once()
+
 
 def test_parse_version_string():
     version_string = "owner/repo@1.0.0"

tests-unit/prompt_server_test/download_models_test.py

@@ -0,0 +1,321 @@
+import pytest
+import aiohttp
+from aiohttp import ClientResponse
+import itertools
+import os 
+from unittest.mock import AsyncMock, patch, MagicMock
+from model_filemanager import download_model, validate_model_subdirectory, track_download_progress, create_model_path, check_file_exists, DownloadStatusType, DownloadModelStatus, validate_filename
+
+class AsyncIteratorMock:
+    """
+    A mock class that simulates an asynchronous iterator.
+    This is used to mimic the behavior of aiohttp's content iterator.
+    """
+    def __init__(self, seq):
+        # Convert the input sequence into an iterator
+        self.iter = iter(seq)
+
+    def __aiter__(self):
+        # This method is called when 'async for' is used
+        return self
+
+    async def __anext__(self):
+        # This method is called for each iteration in an 'async for' loop
+        try:
+            return next(self.iter)
+        except StopIteration:
+            # This is the asynchronous equivalent of StopIteration
+            raise StopAsyncIteration
+
+class ContentMock:
+    """
+    A mock class that simulates the content attribute of an aiohttp ClientResponse.
+    This class provides the iter_chunked method which returns an async iterator of chunks.
+    """
+    def __init__(self, chunks):
+        # Store the chunks that will be returned by the iterator
+        self.chunks = chunks
+
+    def iter_chunked(self, chunk_size):
+        # This method mimics aiohttp's content.iter_chunked()
+        # For simplicity in testing, we ignore chunk_size and just return our predefined chunks
+        return AsyncIteratorMock(self.chunks)
+
+@pytest.mark.asyncio
+async def test_download_model_success():
+    mock_response = AsyncMock(spec=aiohttp.ClientResponse)
+    mock_response.status = 200
+    mock_response.headers = {'Content-Length': '1000'}
+    # Create a mock for content that returns an async iterator directly
+    chunks = [b'a' * 500, b'b' * 300, b'c' * 200]
+    mock_response.content = ContentMock(chunks)
+
+    mock_make_request = AsyncMock(return_value=mock_response)
+    mock_progress_callback = AsyncMock()
+
+    # Mock file operations
+    mock_open = MagicMock()
+    mock_file = MagicMock()
+    mock_open.return_value.__enter__.return_value = mock_file
+    time_values = itertools.count(0, 0.1)
+
+    with patch('model_filemanager.create_model_path', return_value=('models/checkpoints/model.sft', 'checkpoints/model.sft')), \
+         patch('model_filemanager.check_file_exists', return_value=None), \
+         patch('builtins.open', mock_open), \
+         patch('time.time', side_effect=time_values):  # Simulate time passing
+
+        result = await download_model(
+            mock_make_request,
+            'model.sft',
+            'http://example.com/model.sft',
+            'checkpoints',
+            mock_progress_callback
+        )
+
+    # Assert the result
+    assert isinstance(result, DownloadModelStatus)
+    assert result.message == 'Successfully downloaded model.sft'
+    assert result.status == 'completed'
+    assert result.already_existed is False
+
+    # Check progress callback calls
+    assert mock_progress_callback.call_count >= 3  # At least start, one progress update, and completion
+    
+    # Check initial call
+    mock_progress_callback.assert_any_call(
+        'checkpoints/model.sft',
+        DownloadModelStatus(DownloadStatusType.PENDING, 0, "Starting download of model.sft", False)
+    )
+
+    # Check final call
+    mock_progress_callback.assert_any_call(
+        'checkpoints/model.sft',
+        DownloadModelStatus(DownloadStatusType.COMPLETED, 100, "Successfully downloaded model.sft", False)
+    )
+
+    # Verify file writing
+    mock_file.write.assert_any_call(b'a' * 500)
+    mock_file.write.assert_any_call(b'b' * 300)
+    mock_file.write.assert_any_call(b'c' * 200)
+
+    # Verify request was made
+    mock_make_request.assert_called_once_with('http://example.com/model.sft')
+
+@pytest.mark.asyncio
+async def test_download_model_url_request_failure():
+    # Mock dependencies
+    mock_response = AsyncMock(spec=ClientResponse)
+    mock_response.status = 404  # Simulate a "Not Found" error
+    mock_get = AsyncMock(return_value=mock_response)
+    mock_progress_callback = AsyncMock()
+
+    # Mock the create_model_path function
+    with patch('model_filemanager.create_model_path', return_value=('/mock/path/model.safetensors', 'mock/path/model.safetensors')):
+        # Mock the check_file_exists function to return None (file doesn't exist)
+        with patch('model_filemanager.check_file_exists', return_value=None):
+            # Call the function
+            result = await download_model(
+                mock_get,
+                'model.safetensors',
+                'http://example.com/model.safetensors',
+                'mock_directory',
+                mock_progress_callback
+            )
+
+    # Assert the expected behavior
+    assert isinstance(result, DownloadModelStatus)
+    assert result.status == 'error'
+    assert result.message == 'Failed to download model.safetensors. Status code: 404'
+    assert result.already_existed is False
+
+    # Check that progress_callback was called with the correct arguments
+    mock_progress_callback.assert_any_call(
+        'mock_directory/model.safetensors',
+        DownloadModelStatus(
+            status=DownloadStatusType.PENDING,
+            progress_percentage=0,
+            message='Starting download of model.safetensors',
+            already_existed=False
+        )
+    )
+    mock_progress_callback.assert_called_with(
+        'mock_directory/model.safetensors',
+        DownloadModelStatus(
+            status=DownloadStatusType.ERROR,
+            progress_percentage=0,
+            message='Failed to download model.safetensors. Status code: 404',
+            already_existed=False
+        )
+    )
+
+    # Verify that the get method was called with the correct URL
+    mock_get.assert_called_once_with('http://example.com/model.safetensors')
+
+@pytest.mark.asyncio
+async def test_download_model_invalid_model_subdirectory():
+    
+    mock_make_request = AsyncMock()
+    mock_progress_callback = AsyncMock()
+
+    
+    result = await download_model(
+        mock_make_request,
+        'model.sft',
+        'http://example.com/model.sft',
+        '../bad_path',
+        mock_progress_callback
+    )
+
+    # Assert the result
+    assert isinstance(result, DownloadModelStatus)
+    assert result.message == 'Invalid model subdirectory'
+    assert result.status == 'error'
+    assert result.already_existed is False
+
+
+# For create_model_path function
+def test_create_model_path(tmp_path, monkeypatch):
+    mock_models_dir = tmp_path / "models"
+    monkeypatch.setattr('folder_paths.models_dir', str(mock_models_dir))
+    
+    model_name = "test_model.sft"
+    model_directory = "test_dir"
+    
+    file_path, relative_path = create_model_path(model_name, model_directory, mock_models_dir)
+    
+    assert file_path == str(mock_models_dir / model_directory / model_name)
+    assert relative_path == f"{model_directory}/{model_name}"
+    assert os.path.exists(os.path.dirname(file_path))
+
+
+@pytest.mark.asyncio
+async def test_check_file_exists_when_file_exists(tmp_path):
+    file_path = tmp_path / "existing_model.sft"
+    file_path.touch()  # Create an empty file
+    
+    mock_callback = AsyncMock()
+    
+    result = await check_file_exists(str(file_path), "existing_model.sft", mock_callback, "test/existing_model.sft")
+    
+    assert result is not None
+    assert result.status == "completed"
+    assert result.message == "existing_model.sft already exists"
+    assert result.already_existed is True
+    
+    mock_callback.assert_called_once_with(
+        "test/existing_model.sft",
+        DownloadModelStatus(DownloadStatusType.COMPLETED, 100, "existing_model.sft already exists", already_existed=True)
+    )
+
+@pytest.mark.asyncio
+async def test_check_file_exists_when_file_does_not_exist(tmp_path):
+    file_path = tmp_path / "non_existing_model.sft"
+    
+    mock_callback = AsyncMock()
+    
+    result = await check_file_exists(str(file_path), "non_existing_model.sft", mock_callback, "test/non_existing_model.sft")
+    
+    assert result is None
+    mock_callback.assert_not_called()
+
+@pytest.mark.asyncio
+async def test_track_download_progress_no_content_length():
+    mock_response = AsyncMock(spec=aiohttp.ClientResponse)
+    mock_response.headers = {}  # No Content-Length header
+    mock_response.content.iter_chunked.return_value = AsyncIteratorMock([b'a' * 500, b'b' * 500])
+
+    mock_callback = AsyncMock()
+    mock_open = MagicMock(return_value=MagicMock())
+
+    with patch('builtins.open', mock_open):
+        result = await track_download_progress(
+            mock_response, '/mock/path/model.sft', 'model.sft',
+            mock_callback, 'models/model.sft', interval=0.1
+        )
+
+    assert result.status == "completed"
+    # Check that progress was reported even without knowing the total size
+    mock_callback.assert_any_call(
+        'models/model.sft',
+        DownloadModelStatus(DownloadStatusType.IN_PROGRESS, 0, "Downloading model.sft", already_existed=False)
+    )
+
+@pytest.mark.asyncio
+async def test_track_download_progress_interval():
+    mock_response = AsyncMock(spec=aiohttp.ClientResponse)
+    mock_response.headers = {'Content-Length': '1000'}
+    mock_response.content.iter_chunked.return_value = AsyncIteratorMock([b'a' * 100] * 10)
+
+    mock_callback = AsyncMock()
+    mock_open = MagicMock(return_value=MagicMock())
+
+    # Create a mock time function that returns incremental float values
+    mock_time = MagicMock()
+    mock_time.side_effect = [i * 0.5 for i in range(30)]  # This should be enough for 10 chunks
+
+    with patch('builtins.open', mock_open), \
+         patch('time.time', mock_time):
+        await track_download_progress(
+            mock_response, '/mock/path/model.sft', 'model.sft',
+            mock_callback, 'models/model.sft', interval=1.0
+        )
+
+    # Print out the actual call count and the arguments of each call for debugging
+    print(f"mock_callback was called {mock_callback.call_count} times")
+    for i, call in enumerate(mock_callback.call_args_list):
+        args, kwargs = call
+        print(f"Call {i + 1}: {args[1].status}, Progress: {args[1].progress_percentage:.2f}%")
+
+    # Assert that progress was updated at least 3 times (start, at least one interval, and end)
+    assert mock_callback.call_count >= 3, f"Expected at least 3 calls, but got {mock_callback.call_count}"
+
+    # Verify the first and last calls
+    first_call = mock_callback.call_args_list[0]
+    assert first_call[0][1].status == "in_progress"
+    # Allow for some initial progress, but it should be less than 50%
+    assert 0 <= first_call[0][1].progress_percentage < 50, f"First call progress was {first_call[0][1].progress_percentage}%"
+
+    last_call = mock_callback.call_args_list[-1]
+    assert last_call[0][1].status == "completed"
+    assert last_call[0][1].progress_percentage == 100
+
+def test_valid_subdirectory():
+    assert validate_model_subdirectory("valid-model123") is True
+
+def test_subdirectory_too_long():
+    assert validate_model_subdirectory("a" * 51) is False
+
+def test_subdirectory_with_double_dots():
+    assert validate_model_subdirectory("model/../unsafe") is False
+
+def test_subdirectory_with_slash():
+    assert validate_model_subdirectory("model/unsafe") is False
+
+def test_subdirectory_with_special_characters():
+    assert validate_model_subdirectory("model@unsafe") is False
+
+def test_subdirectory_with_underscore_and_dash():
+    assert validate_model_subdirectory("valid_model-name") is True
+
+def test_empty_subdirectory():
+    assert validate_model_subdirectory("") is False
+
+@pytest.mark.parametrize("filename, expected", [
+    ("valid_model.safetensors", True),
+    ("valid_model.sft", True),
+    ("valid model.safetensors", True), # Test with space
+    ("UPPERCASE_MODEL.SAFETENSORS", True),
+    ("model_with.multiple.dots.pt", False),
+    ("", False),  # Empty string
+    ("../../../etc/passwd", False),  # Path traversal attempt
+    ("/etc/passwd", False),  # Absolute path
+    ("\\windows\\system32\\config\\sam", False),  # Windows path
+    (".hidden_file.pt", False),  # Hidden file
+    ("invalid<char>.ckpt", False),  # Invalid character
+    ("invalid?.ckpt", False),  # Another invalid character
+    ("very" * 100 + ".safetensors", False),  # Too long filename
+    ("\nmodel_with_newline.pt", False),  # Newline character
+    ("model_with_emoji😊.pt", False),  # Emoji in filename
+])
+def test_validate_filename(filename, expected):
+    assert validate_filename(filename) == expected

tests-unit/requirements.txt

@@ -1 +1,3 @@
 pytest>=7.8.0
+pytest-aiohttp
+pytest-asyncio

tests-unit/server/routes/internal_routes_test.py

@@ -0,0 +1,115 @@
+import pytest
+from aiohttp import web
+from unittest.mock import MagicMock, patch
+from api_server.routes.internal.internal_routes import InternalRoutes
+from api_server.services.file_service import FileService
+from folder_paths import models_dir, user_directory, output_directory
+
+
+@pytest.fixture
+def internal_routes():
+    return InternalRoutes()
+
+@pytest.fixture
+def aiohttp_client_factory(aiohttp_client, internal_routes):
+    async def _get_client():
+        app = internal_routes.get_app()
+        return await aiohttp_client(app)
+    return _get_client
+
+@pytest.mark.asyncio
+async def test_list_files_valid_directory(aiohttp_client_factory, internal_routes):
+    mock_file_list = [
+        {"name": "file1.txt", "path": "file1.txt", "type": "file", "size": 100},
+        {"name": "dir1", "path": "dir1", "type": "directory"}
+    ]
+    internal_routes.file_service.list_files = MagicMock(return_value=mock_file_list)
+    client = await aiohttp_client_factory()
+    resp = await client.get('/files?directory=models')
+    assert resp.status == 200
+    data = await resp.json()
+    assert 'files' in data
+    assert len(data['files']) == 2
+    assert data['files'] == mock_file_list
+
+    # Check other valid directories
+    resp = await client.get('/files?directory=user')
+    assert resp.status == 200
+    resp = await client.get('/files?directory=output')
+    assert resp.status == 200
+
+@pytest.mark.asyncio
+async def test_list_files_invalid_directory(aiohttp_client_factory, internal_routes):
+    internal_routes.file_service.list_files = MagicMock(side_effect=ValueError("Invalid directory key"))
+    client = await aiohttp_client_factory()
+    resp = await client.get('/files?directory=invalid')
+    assert resp.status == 400
+    data = await resp.json()
+    assert 'error' in data
+    assert data['error'] == "Invalid directory key"
+
+@pytest.mark.asyncio
+async def test_list_files_exception(aiohttp_client_factory, internal_routes):
+    internal_routes.file_service.list_files = MagicMock(side_effect=Exception("Unexpected error"))
+    client = await aiohttp_client_factory()
+    resp = await client.get('/files?directory=models')
+    assert resp.status == 500
+    data = await resp.json()
+    assert 'error' in data
+    assert data['error'] == "Unexpected error"
+
+@pytest.mark.asyncio
+async def test_list_files_no_directory_param(aiohttp_client_factory, internal_routes):
+    mock_file_list = []
+    internal_routes.file_service.list_files = MagicMock(return_value=mock_file_list)
+    client = await aiohttp_client_factory()
+    resp = await client.get('/files')
+    assert resp.status == 200
+    data = await resp.json()
+    assert 'files' in data
+    assert len(data['files']) == 0
+
+def test_setup_routes(internal_routes):
+    internal_routes.setup_routes()
+    routes = internal_routes.routes
+    assert any(route.method == 'GET' and str(route.path) == '/files' for route in routes)
+
+def test_get_app(internal_routes):
+    app = internal_routes.get_app()
+    assert isinstance(app, web.Application)
+    assert internal_routes._app is not None
+
+def test_get_app_reuse(internal_routes):
+    app1 = internal_routes.get_app()
+    app2 = internal_routes.get_app()
+    assert app1 is app2
+
+@pytest.mark.asyncio
+async def test_routes_added_to_app(aiohttp_client_factory, internal_routes):
+    client = await aiohttp_client_factory()
+    try:
+        resp = await client.get('/files')
+        print(f"Response received: status {resp.status}")
+    except Exception as e:
+        print(f"Exception occurred during GET request: {e}")
+        raise
+
+    assert resp.status != 404, "Route /files does not exist"
+
+@pytest.mark.asyncio
+async def test_file_service_initialization():
+    with patch('api_server.routes.internal.internal_routes.FileService') as MockFileService:
+        # Create a mock instance
+        mock_file_service_instance = MagicMock(spec=FileService)
+        MockFileService.return_value = mock_file_service_instance
+        internal_routes = InternalRoutes()
+
+        # Check if FileService was initialized with the correct parameters
+        MockFileService.assert_called_once_with({
+            "models": models_dir,
+            "user": user_directory,
+            "output": output_directory
+        })
+
+        # Verify that the file_service attribute of InternalRoutes is set
+        assert internal_routes.file_service == mock_file_service_instance

tests-unit/server/services/file_service_test.py

@@ -0,0 +1,54 @@
+import pytest
+from unittest.mock import MagicMock
+from api_server.services.file_service import FileService
+
+@pytest.fixture
+def mock_file_system_ops():
+    return MagicMock()
+
+@pytest.fixture
+def file_service(mock_file_system_ops):
+    allowed_directories = {
+        "models": "/path/to/models",
+        "user": "/path/to/user",
+        "output": "/path/to/output"
+    }
+    return FileService(allowed_directories, file_system_ops=mock_file_system_ops)
+
+def test_list_files_valid_directory(file_service, mock_file_system_ops):
+    mock_file_system_ops.walk_directory.return_value = [
+        {"name": "file1.txt", "path": "file1.txt", "type": "file", "size": 100},
+        {"name": "dir1", "path": "dir1", "type": "directory"}
+    ]
+    
+    result = file_service.list_files("models")
+    
+    assert len(result) == 2
+    assert result[0]["name"] == "file1.txt"
+    assert result[1]["name"] == "dir1"
+    mock_file_system_ops.walk_directory.assert_called_once_with("/path/to/models")
+
+def test_list_files_invalid_directory(file_service):
+    # Does not support walking directories outside of the allowed directories
+    with pytest.raises(ValueError, match="Invalid directory key"):
+        file_service.list_files("invalid_key")
+
+def test_list_files_empty_directory(file_service, mock_file_system_ops):
+    mock_file_system_ops.walk_directory.return_value = []
+    
+    result = file_service.list_files("models")
+    
+    assert len(result) == 0
+    mock_file_system_ops.walk_directory.assert_called_once_with("/path/to/models")
+
+@pytest.mark.parametrize("directory_key", ["models", "user", "output"])
+def test_list_files_all_allowed_directories(file_service, mock_file_system_ops, directory_key):
+    mock_file_system_ops.walk_directory.return_value = [
+        {"name": f"file_{directory_key}.txt", "path": f"file_{directory_key}.txt", "type": "file", "size": 100}
+    ]
+    
+    result = file_service.list_files(directory_key)
+    
+    assert len(result) == 1
+    assert result[0]["name"] == f"file_{directory_key}.txt"
+    mock_file_system_ops.walk_directory.assert_called_once_with(f"/path/to/{directory_key}")

tests-unit/server/utils/file_operations_test.py

@@ -0,0 +1,42 @@
+import pytest
+from typing import List
+from api_server.utils.file_operations import FileSystemOperations, FileSystemItem, is_file_info
+
+@pytest.fixture
+def temp_directory(tmp_path):
+    # Create a temporary directory structure
+    dir1 = tmp_path / "dir1"
+    dir2 = tmp_path / "dir2"
+    dir1.mkdir()
+    dir2.mkdir()
+    (dir1 / "file1.txt").write_text("content1")
+    (dir2 / "file2.txt").write_text("content2")
+    (tmp_path / "file3.txt").write_text("content3")
+    return tmp_path
+
+def test_walk_directory(temp_directory):
+    result: List[FileSystemItem] = FileSystemOperations.walk_directory(str(temp_directory))
+    
+    assert len(result) == 5  # 2 directories and 3 files
+    
+    files = [item for item in result if item['type'] == 'file']
+    dirs = [item for item in result if item['type'] == 'directory']
+    
+    assert len(files) == 3
+    assert len(dirs) == 2
+    
+    file_names = {file['name'] for file in files}
+    assert file_names == {'file1.txt', 'file2.txt', 'file3.txt'}
+    
+    dir_names = {dir['name'] for dir in dirs}
+    assert dir_names == {'dir1', 'dir2'}
+
+def test_walk_directory_empty(tmp_path):
+    result = FileSystemOperations.walk_directory(str(tmp_path))
+    assert len(result) == 0
+
+def test_walk_directory_file_size(temp_directory):
+    result: List[FileSystemItem] = FileSystemOperations.walk_directory(str(temp_directory))
+    files = [item for item in result if is_file_info(item)]
+    for file in files:
+        assert file['size'] > 0  # Assuming all files have some content

tests/inference/extra_model_paths.yaml

@@ -0,0 +1,4 @@
+# Config for testing nodes
+testing:
+    custom_nodes: tests/inference/testing_nodes
+

tests/inference/test_execution.py

@@ -0,0 +1,499 @@
+from io import BytesIO
+import numpy
+from PIL import Image
+import pytest
+from pytest import fixture
+import time
+import torch
+from typing import Union, Dict
+import json
+import subprocess
+import websocket #NOTE: websocket-client (https://github.com/websocket-client/websocket-client)
+import uuid
+import urllib.request
+import urllib.parse
+import urllib.error
+from comfy_execution.graph_utils import GraphBuilder, Node
+
+class RunResult:
+    def __init__(self, prompt_id: str):
+        self.outputs: Dict[str,Dict] = {}
+        self.runs: Dict[str,bool] = {}
+        self.prompt_id: str = prompt_id
+
+    def get_output(self, node: Node):
+        return self.outputs.get(node.id, None)
+
+    def did_run(self, node: Node):
+        return self.runs.get(node.id, False)
+
+    def get_images(self, node: Node):
+        output = self.get_output(node)
+        if output is None:
+            return []
+        return output.get('image_objects', [])
+
+    def get_prompt_id(self):
+        return self.prompt_id
+
+class ComfyClient:
+    def __init__(self):
+        self.test_name = ""
+
+    def connect(self, 
+                    listen:str = '127.0.0.1', 
+                    port:Union[str,int] = 8188,
+                    client_id: str = str(uuid.uuid4())
+                    ):
+        self.client_id = client_id
+        self.server_address = f"{listen}:{port}"
+        ws = websocket.WebSocket()
+        ws.connect("ws://{}/ws?clientId={}".format(self.server_address, self.client_id))
+        self.ws = ws
+
+    def queue_prompt(self, prompt):
+        p = {"prompt": prompt, "client_id": self.client_id}
+        data = json.dumps(p).encode('utf-8')
+        req =  urllib.request.Request("http://{}/prompt".format(self.server_address), data=data)
+        return json.loads(urllib.request.urlopen(req).read())
+
+    def get_image(self, filename, subfolder, folder_type):
+        data = {"filename": filename, "subfolder": subfolder, "type": folder_type}
+        url_values = urllib.parse.urlencode(data)
+        with urllib.request.urlopen("http://{}/view?{}".format(self.server_address, url_values)) as response:
+            return response.read()
+
+    def get_history(self, prompt_id):
+        with urllib.request.urlopen("http://{}/history/{}".format(self.server_address, prompt_id)) as response:
+            return json.loads(response.read())
+
+    def set_test_name(self, name):
+        self.test_name = name
+
+    def run(self, graph):
+        prompt = graph.finalize()
+        for node in graph.nodes.values():
+            if node.class_type == 'SaveImage':
+                node.inputs['filename_prefix'] = self.test_name
+
+        prompt_id = self.queue_prompt(prompt)['prompt_id']
+        result = RunResult(prompt_id)
+        while True:
+            out = self.ws.recv()
+            if isinstance(out, str):
+                message = json.loads(out)
+                if message['type'] == 'executing':
+                    data = message['data']
+                    if data['prompt_id'] != prompt_id:
+                        continue
+                    if data['node'] is None:
+                        break
+                    result.runs[data['node']] = True
+                elif message['type'] == 'execution_error':
+                    raise Exception(message['data'])
+                elif message['type'] == 'execution_cached':
+                    pass # Probably want to store this off for testing
+
+        history = self.get_history(prompt_id)[prompt_id]
+        for o in history['outputs']:
+            for node_id in history['outputs']:
+                node_output = history['outputs'][node_id]
+                result.outputs[node_id] = node_output
+                if 'images' in node_output:
+                    images_output = []
+                    for image in node_output['images']:
+                        image_data = self.get_image(image['filename'], image['subfolder'], image['type'])
+                        image_obj = Image.open(BytesIO(image_data))
+                        images_output.append(image_obj)
+                    node_output['image_objects'] = images_output
+
+        return result
+
+#
+# Loop through these variables
+#
+@pytest.mark.execution
+class TestExecution:
+    #
+    # Initialize server and client
+    #
+    @fixture(scope="class", autouse=True, params=[
+        # (use_lru, lru_size)
+        (False, 0),
+        (True, 0),
+        (True, 100),
+    ])
+    def _server(self, args_pytest, request):
+        # Start server
+        pargs = [
+            'python','main.py', 
+            '--output-directory', args_pytest["output_dir"],
+            '--listen', args_pytest["listen"],
+            '--port', str(args_pytest["port"]),
+            '--extra-model-paths-config', 'tests/inference/extra_model_paths.yaml',
+        ]
+        use_lru, lru_size = request.param
+        if use_lru:
+            pargs += ['--cache-lru', str(lru_size)]
+        print("Running server with args:", pargs)
+        p = subprocess.Popen(pargs)
+        yield
+        p.kill()
+        torch.cuda.empty_cache()
+
+    def start_client(self, listen:str, port:int):
+        # Start client
+        comfy_client = ComfyClient()
+        # Connect to server (with retries)
+        n_tries = 5
+        for i in range(n_tries):
+            time.sleep(4)
+            try:
+                comfy_client.connect(listen=listen, port=port)
+            except ConnectionRefusedError as e:
+                print(e)
+                print(f"({i+1}/{n_tries}) Retrying...")
+            else:
+                break
+        return comfy_client
+
+    @fixture(scope="class", autouse=True)
+    def shared_client(self, args_pytest, _server):
+        client = self.start_client(args_pytest["listen"], args_pytest["port"])
+        yield client
+        del client
+        torch.cuda.empty_cache()
+
+    @fixture
+    def client(self, shared_client, request):
+        shared_client.set_test_name(f"execution[{request.node.name}]")
+        yield shared_client
+
+    @fixture
+    def builder(self, request):
+        yield GraphBuilder(prefix=request.node.name)
+
+    def test_lazy_input(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        mask = g.node("StubMask", value=0.0, height=512, width=512, batch_size=1)
+
+        lazy_mix = g.node("TestLazyMixImages", image1=input1.out(0), image2=input2.out(0), mask=mask.out(0))
+        output = g.node("SaveImage", images=lazy_mix.out(0))
+        result = client.run(g)
+
+        result_image = result.get_images(output)[0]
+        assert numpy.array(result_image).any() == 0, "Image should be black"
+        assert result.did_run(input1)
+        assert not result.did_run(input2)
+        assert result.did_run(mask)
+        assert result.did_run(lazy_mix)
+
+    def test_full_cache(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input2 = g.node("StubImage", content="NOISE", height=512, width=512, batch_size=1)
+        mask = g.node("StubMask", value=0.5, height=512, width=512, batch_size=1)
+
+        lazy_mix = g.node("TestLazyMixImages", image1=input1.out(0), image2=input2.out(0), mask=mask.out(0))
+        g.node("SaveImage", images=lazy_mix.out(0))
+
+        client.run(g)
+        result2 = client.run(g)
+        for node_id, node in g.nodes.items():
+            assert not result2.did_run(node), f"Node {node_id} ran, but should have been cached"
+
+    def test_partial_cache(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input2 = g.node("StubImage", content="NOISE", height=512, width=512, batch_size=1)
+        mask = g.node("StubMask", value=0.5, height=512, width=512, batch_size=1)
+
+        lazy_mix = g.node("TestLazyMixImages", image1=input1.out(0), image2=input2.out(0), mask=mask.out(0))
+        g.node("SaveImage", images=lazy_mix.out(0))
+
+        client.run(g)
+        mask.inputs['value'] = 0.4
+        result2 = client.run(g)
+        assert not result2.did_run(input1), "Input1 should have been cached"
+        assert not result2.did_run(input2), "Input2 should have been cached"
+
+    def test_error(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        # Different size of the two images
+        input2 = g.node("StubImage", content="NOISE", height=256, width=256, batch_size=1)
+        mask = g.node("StubMask", value=0.5, height=512, width=512, batch_size=1)
+
+        lazy_mix = g.node("TestLazyMixImages", image1=input1.out(0), image2=input2.out(0), mask=mask.out(0))
+        g.node("SaveImage", images=lazy_mix.out(0))
+
+        try:
+            client.run(g)
+            assert False, "Should have raised an error"
+        except Exception as e:
+            assert 'prompt_id' in e.args[0], f"Did not get back a proper error message: {e}"
+
+    @pytest.mark.parametrize("test_value, expect_error", [
+        (5, True),
+        ("foo", True),
+        (5.0, False),
+    ])
+    def test_validation_error_literal(self, test_value, expect_error, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        validation1 = g.node("TestCustomValidation1", input1=test_value, input2=3.0)
+        g.node("SaveImage", images=validation1.out(0))
+
+        if expect_error:
+            with pytest.raises(urllib.error.HTTPError):
+                client.run(g)
+        else:
+            client.run(g)
+
+    @pytest.mark.parametrize("test_type, test_value", [
+        ("StubInt", 5),
+        ("StubFloat", 5.0)
+    ])
+    def test_validation_error_edge1(self, test_type, test_value, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        stub = g.node(test_type, value=test_value)
+        validation1 = g.node("TestCustomValidation1", input1=stub.out(0), input2=3.0)
+        g.node("SaveImage", images=validation1.out(0))
+
+        with pytest.raises(urllib.error.HTTPError):
+            client.run(g)
+
+    @pytest.mark.parametrize("test_type, test_value, expect_error", [
+        ("StubInt", 5, True),
+        ("StubFloat", 5.0, False)
+    ])
+    def test_validation_error_edge2(self, test_type, test_value, expect_error, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        stub = g.node(test_type, value=test_value)
+        validation2 = g.node("TestCustomValidation2", input1=stub.out(0), input2=3.0)
+        g.node("SaveImage", images=validation2.out(0))
+
+        if expect_error:
+            with pytest.raises(urllib.error.HTTPError):
+                client.run(g)
+        else:
+            client.run(g)
+
+    @pytest.mark.parametrize("test_type, test_value, expect_error", [
+        ("StubInt", 5, True),
+        ("StubFloat", 5.0, False)
+    ])
+    def test_validation_error_edge3(self, test_type, test_value, expect_error, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        stub = g.node(test_type, value=test_value)
+        validation3 = g.node("TestCustomValidation3", input1=stub.out(0), input2=3.0)
+        g.node("SaveImage", images=validation3.out(0))
+
+        if expect_error:
+            with pytest.raises(urllib.error.HTTPError):
+                client.run(g)
+        else:
+            client.run(g)
+
+    @pytest.mark.parametrize("test_type, test_value, expect_error", [
+        ("StubInt", 5, True),
+        ("StubFloat", 5.0, False)
+    ])
+    def test_validation_error_edge4(self, test_type, test_value, expect_error, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        stub = g.node(test_type, value=test_value)
+        validation4 = g.node("TestCustomValidation4", input1=stub.out(0), input2=3.0)
+        g.node("SaveImage", images=validation4.out(0))
+
+        if expect_error:
+            with pytest.raises(urllib.error.HTTPError):
+                client.run(g)
+        else:
+            client.run(g)
+
+    @pytest.mark.parametrize("test_value1, test_value2, expect_error", [
+        (0.0, 0.5, False),
+        (0.0, 5.0, False),
+        (0.0, 7.0, True)
+    ])
+    def test_validation_error_kwargs(self, test_value1, test_value2, expect_error, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        validation5 = g.node("TestCustomValidation5", input1=test_value1, input2=test_value2)
+        g.node("SaveImage", images=validation5.out(0))
+
+        if expect_error:
+            with pytest.raises(urllib.error.HTTPError):
+                client.run(g)
+        else:
+            client.run(g)
+
+    def test_cycle_error(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        mask = g.node("StubMask", value=0.5, height=512, width=512, batch_size=1)
+
+        lazy_mix1 = g.node("TestLazyMixImages", image1=input1.out(0), mask=mask.out(0))
+        lazy_mix2 = g.node("TestLazyMixImages", image1=lazy_mix1.out(0), image2=input2.out(0), mask=mask.out(0))
+        g.node("SaveImage", images=lazy_mix2.out(0))
+
+        # When the cycle exists on initial submission, it should raise a validation error
+        with pytest.raises(urllib.error.HTTPError):
+            client.run(g)
+
+    def test_dynamic_cycle_error(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        generator = g.node("TestDynamicDependencyCycle", input1=input1.out(0), input2=input2.out(0))
+        g.node("SaveImage", images=generator.out(0))
+
+        # When the cycle is in a graph that is generated dynamically, it should raise a runtime error
+        try:
+            client.run(g)
+            assert False, "Should have raised an error"
+        except Exception as e:
+            assert 'prompt_id' in e.args[0], f"Did not get back a proper error message: {e}"
+            assert e.args[0]['node_id'] == generator.id, "Error should have been on the generator node"
+
+    def test_missing_node_error(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input2 = g.node("StubImage", id="removeme", content="WHITE", height=512, width=512, batch_size=1)
+        input3 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        mask = g.node("StubMask", value=0.5, height=512, width=512, batch_size=1)
+        mix1 = g.node("TestLazyMixImages", image1=input1.out(0), image2=input2.out(0), mask=mask.out(0))
+        mix2 = g.node("TestLazyMixImages", image1=input1.out(0), image2=input3.out(0), mask=mask.out(0))
+        # We have multiple outputs. The first is invalid, but the second is valid
+        g.node("SaveImage", images=mix1.out(0))
+        g.node("SaveImage", images=mix2.out(0))
+        g.remove_node("removeme")
+        
+        client.run(g)
+
+        # Add back in the missing node to make sure the error doesn't break the server
+        input2 = g.node("StubImage", id="removeme", content="WHITE", height=512, width=512, batch_size=1)
+        client.run(g)
+
+    def test_custom_is_changed(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        # Creating the nodes in this specific order previously caused a bug
+        save = g.node("SaveImage")
+        is_changed = g.node("TestCustomIsChanged", should_change=False)
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+
+        save.set_input('images', is_changed.out(0))
+        is_changed.set_input('image', input1.out(0))
+
+        result1 = client.run(g)
+        result2 = client.run(g)
+        is_changed.set_input('should_change', True)
+        result3 = client.run(g)
+        result4 = client.run(g)
+        assert result1.did_run(is_changed), "is_changed should have been run"
+        assert not result2.did_run(is_changed), "is_changed should have been cached"
+        assert result3.did_run(is_changed), "is_changed should have been re-run"
+        assert result4.did_run(is_changed), "is_changed should not have been cached"
+
+    def test_undeclared_inputs(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        input3 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input4 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        average = g.node("TestVariadicAverage", input1=input1.out(0), input2=input2.out(0), input3=input3.out(0), input4=input4.out(0))
+        output = g.node("SaveImage", images=average.out(0))
+
+        result = client.run(g)
+        result_image = result.get_images(output)[0]
+        expected = 255 // 4
+        assert numpy.array(result_image).min() == expected and numpy.array(result_image).max() == expected, "Image should be grey"
+
+    def test_for_loop(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        iterations = 4
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        is_changed = g.node("TestCustomIsChanged", should_change=True, image=input2.out(0))
+        for_open = g.node("TestForLoopOpen", remaining=iterations, initial_value1=is_changed.out(0))
+        average = g.node("TestVariadicAverage", input1=input1.out(0), input2=for_open.out(2))
+        for_close = g.node("TestForLoopClose", flow_control=for_open.out(0), initial_value1=average.out(0))
+        output = g.node("SaveImage", images=for_close.out(0))
+
+        for iterations in range(1, 5):
+            for_open.set_input('remaining', iterations)
+            result = client.run(g)
+            result_image = result.get_images(output)[0]
+            expected = 255 // (2 ** iterations)
+            assert numpy.array(result_image).min() == expected and numpy.array(result_image).max() == expected, "Image should be grey"
+            assert result.did_run(is_changed)
+
+    def test_mixed_expansion_returns(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        val_list = g.node("TestMakeListNode", value1=0.1, value2=0.2, value3=0.3)
+        mixed = g.node("TestMixedExpansionReturns", input1=val_list.out(0))
+        output_dynamic = g.node("SaveImage", images=mixed.out(0))
+        output_literal = g.node("SaveImage", images=mixed.out(1))
+
+        result = client.run(g)
+        images_dynamic = result.get_images(output_dynamic)
+        assert len(images_dynamic) == 3, "Should have 2 images"
+        assert numpy.array(images_dynamic[0]).min() == 25 and numpy.array(images_dynamic[0]).max() == 25, "First image should be 0.1"
+        assert numpy.array(images_dynamic[1]).min() == 51 and numpy.array(images_dynamic[1]).max() == 51, "Second image should be 0.2"
+        assert numpy.array(images_dynamic[2]).min() == 76 and numpy.array(images_dynamic[2]).max() == 76, "Third image should be 0.3"
+
+        images_literal = result.get_images(output_literal)
+        assert len(images_literal) == 3, "Should have 2 images"
+        for i in range(3):
+            assert numpy.array(images_literal[i]).min() == 255 and numpy.array(images_literal[i]).max() == 255, "All images should be white"
+
+    def test_mixed_lazy_results(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        val_list = g.node("TestMakeListNode", value1=0.0, value2=0.5, value3=1.0)
+        mask = g.node("StubMask", value=val_list.out(0), height=512, width=512, batch_size=1)
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        mix = g.node("TestLazyMixImages", image1=input1.out(0), image2=input2.out(0), mask=mask.out(0))
+        rebatch = g.node("RebatchImages", images=mix.out(0), batch_size=3)
+        output = g.node("SaveImage", images=rebatch.out(0))
+
+        result = client.run(g)
+        images = result.get_images(output)
+        assert len(images) == 3, "Should have 3 image"
+        assert numpy.array(images[0]).min() == 0 and numpy.array(images[0]).max() == 0, "First image should be 0.0"
+        assert numpy.array(images[1]).min() == 127 and numpy.array(images[1]).max() == 127, "Second image should be 0.5"
+        assert numpy.array(images[2]).min() == 255 and numpy.array(images[2]).max() == 255, "Third image should be 1.0"
+
+    def test_output_reuse(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+
+        output1 = g.node("SaveImage", images=input1.out(0))
+        output2 = g.node("SaveImage", images=input1.out(0))
+
+        result = client.run(g)
+        images1 = result.get_images(output1)
+        images2 = result.get_images(output2)
+        assert len(images1) == 1, "Should have 1 image"
+        assert len(images2) == 1, "Should have 1 image"
+
+
+    # This tests that only constant outputs are used in the call to `IS_CHANGED`
+    def test_is_changed_with_outputs(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        input1 = g.node("StubConstantImage", value=0.5, height=512, width=512, batch_size=1)
+        test_node = g.node("TestIsChangedWithConstants", image=input1.out(0), value=0.5)
+
+        output = g.node("PreviewImage", images=test_node.out(0))
+
+        result = client.run(g)
+        images = result.get_images(output)
+        assert len(images) == 1, "Should have 1 image"
+        assert numpy.array(images[0]).min() == 63 and numpy.array(images[0]).max() == 63, "Image should have value 0.25"
+
+        result = client.run(g)
+        images = result.get_images(output)
+        assert len(images) == 1, "Should have 1 image"
+        assert numpy.array(images[0]).min() == 63 and numpy.array(images[0]).max() == 63, "Image should have value 0.25"
+        assert not result.did_run(test_node), "The execution should have been cached"

tests/inference/testing_nodes/testing-pack/__init__.py

@@ -0,0 +1,23 @@
+from .specific_tests import TEST_NODE_CLASS_MAPPINGS, TEST_NODE_DISPLAY_NAME_MAPPINGS
+from .flow_control import FLOW_CONTROL_NODE_CLASS_MAPPINGS, FLOW_CONTROL_NODE_DISPLAY_NAME_MAPPINGS
+from .util import UTILITY_NODE_CLASS_MAPPINGS, UTILITY_NODE_DISPLAY_NAME_MAPPINGS
+from .conditions import CONDITION_NODE_CLASS_MAPPINGS, CONDITION_NODE_DISPLAY_NAME_MAPPINGS
+from .stubs import TEST_STUB_NODE_CLASS_MAPPINGS, TEST_STUB_NODE_DISPLAY_NAME_MAPPINGS
+
+# NODE_CLASS_MAPPINGS = GENERAL_NODE_CLASS_MAPPINGS.update(COMPONENT_NODE_CLASS_MAPPINGS)
+# NODE_DISPLAY_NAME_MAPPINGS = GENERAL_NODE_DISPLAY_NAME_MAPPINGS.update(COMPONENT_NODE_DISPLAY_NAME_MAPPINGS)
+
+NODE_CLASS_MAPPINGS = {}
+NODE_CLASS_MAPPINGS.update(TEST_NODE_CLASS_MAPPINGS)
+NODE_CLASS_MAPPINGS.update(FLOW_CONTROL_NODE_CLASS_MAPPINGS)
+NODE_CLASS_MAPPINGS.update(UTILITY_NODE_CLASS_MAPPINGS)
+NODE_CLASS_MAPPINGS.update(CONDITION_NODE_CLASS_MAPPINGS)
+NODE_CLASS_MAPPINGS.update(TEST_STUB_NODE_CLASS_MAPPINGS)
+
+NODE_DISPLAY_NAME_MAPPINGS = {}
+NODE_DISPLAY_NAME_MAPPINGS.update(TEST_NODE_DISPLAY_NAME_MAPPINGS)
+NODE_DISPLAY_NAME_MAPPINGS.update(FLOW_CONTROL_NODE_DISPLAY_NAME_MAPPINGS)
+NODE_DISPLAY_NAME_MAPPINGS.update(UTILITY_NODE_DISPLAY_NAME_MAPPINGS)
+NODE_DISPLAY_NAME_MAPPINGS.update(CONDITION_NODE_DISPLAY_NAME_MAPPINGS)
+NODE_DISPLAY_NAME_MAPPINGS.update(TEST_STUB_NODE_DISPLAY_NAME_MAPPINGS)
+

tests/inference/testing_nodes/testing-pack/conditions.py

@@ -0,0 +1,194 @@
+import re
+import torch
+
+class TestIntConditions:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "a": ("INT", {"default": 0, "min": -0xffffffffffffffff, "max": 0xffffffffffffffff, "step": 1}),
+                "b": ("INT", {"default": 0, "min": -0xffffffffffffffff, "max": 0xffffffffffffffff, "step": 1}),
+                "operation": (["==", "!=", "<", ">", "<=", ">="],),
+            },
+        }
+
+    RETURN_TYPES = ("BOOLEAN",)
+    FUNCTION = "int_condition"
+
+    CATEGORY = "Testing/Logic"
+
+    def int_condition(self, a, b, operation):
+        if operation == "==":
+            return (a == b,)
+        elif operation == "!=":
+            return (a != b,)
+        elif operation == "<":
+            return (a < b,)
+        elif operation == ">":
+            return (a > b,)
+        elif operation == "<=":
+            return (a <= b,)
+        elif operation == ">=":
+            return (a >= b,)
+
+
+class TestFloatConditions:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "a": ("FLOAT", {"default": 0, "min": -999999999999.0, "max": 999999999999.0, "step": 1}),
+                "b": ("FLOAT", {"default": 0, "min": -999999999999.0, "max": 999999999999.0, "step": 1}),
+                "operation": (["==", "!=", "<", ">", "<=", ">="],),
+            },
+        }
+
+    RETURN_TYPES = ("BOOLEAN",)
+    FUNCTION = "float_condition"
+
+    CATEGORY = "Testing/Logic"
+
+    def float_condition(self, a, b, operation):
+        if operation == "==":
+            return (a == b,)
+        elif operation == "!=":
+            return (a != b,)
+        elif operation == "<":
+            return (a < b,)
+        elif operation == ">":
+            return (a > b,)
+        elif operation == "<=":
+            return (a <= b,)
+        elif operation == ">=":
+            return (a >= b,)
+
+class TestStringConditions:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "a": ("STRING", {"multiline": False}),
+                "b": ("STRING", {"multiline": False}),
+                "operation": (["a == b", "a != b", "a IN b", "a MATCH REGEX(b)", "a BEGINSWITH b", "a ENDSWITH b"],),
+                "case_sensitive": ("BOOLEAN", {"default": True}),
+            },
+        }
+
+    RETURN_TYPES = ("BOOLEAN",)
+    FUNCTION = "string_condition"
+
+    CATEGORY = "Testing/Logic"
+
+    def string_condition(self, a, b, operation, case_sensitive):
+        if not case_sensitive:
+            a = a.lower()
+            b = b.lower()
+
+        if operation == "a == b":
+            return (a == b,)
+        elif operation == "a != b":
+            return (a != b,)
+        elif operation == "a IN b":
+            return (a in b,)
+        elif operation == "a MATCH REGEX(b)":
+            try:
+                return (re.match(b, a) is not None,)
+            except:
+                return (False,)
+        elif operation == "a BEGINSWITH b":
+            return (a.startswith(b),)
+        elif operation == "a ENDSWITH b":
+            return (a.endswith(b),)
+
+class TestToBoolNode:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": ("*",),
+            },
+            "optional": {
+                "invert": ("BOOLEAN", {"default": False}),
+            },
+        }
+
+    RETURN_TYPES = ("BOOLEAN",)
+    FUNCTION = "to_bool"
+
+    CATEGORY = "Testing/Logic"
+
+    def to_bool(self, value, invert = False):
+        if isinstance(value, torch.Tensor):
+            if value.max().item() == 0 and value.min().item() == 0:
+                result = False
+            else:
+                result = True
+        else:
+            try:
+                result = bool(value)
+            except:
+                # Can't convert it? Well then it's something or other. I dunno, I'm not a Python programmer.
+                result = True
+
+        if invert:
+            result = not result
+
+        return (result,)
+
+class TestBoolOperationNode:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "a": ("BOOLEAN",),
+                "b": ("BOOLEAN",),
+                "op": (["a AND b", "a OR b", "a XOR b", "NOT a"],),
+            },
+        }
+
+    RETURN_TYPES = ("BOOLEAN",)
+    FUNCTION = "bool_operation"
+
+    CATEGORY = "Testing/Logic"
+
+    def bool_operation(self, a, b, op):
+        if op == "a AND b":
+            return (a and b,)
+        elif op == "a OR b":
+            return (a or b,)
+        elif op == "a XOR b":
+            return (a ^ b,)
+        elif op == "NOT a":
+            return (not a,)
+
+
+CONDITION_NODE_CLASS_MAPPINGS = {
+    "TestIntConditions": TestIntConditions,
+    "TestFloatConditions": TestFloatConditions,
+    "TestStringConditions": TestStringConditions,
+    "TestToBoolNode": TestToBoolNode,
+    "TestBoolOperationNode": TestBoolOperationNode,
+}
+
+CONDITION_NODE_DISPLAY_NAME_MAPPINGS = {
+    "TestIntConditions": "Int Condition",
+    "TestFloatConditions": "Float Condition",
+    "TestStringConditions": "String Condition",
+    "TestToBoolNode": "To Bool",
+    "TestBoolOperationNode": "Bool Operation",
+}

tests/inference/testing_nodes/testing-pack/flow_control.py

@@ -0,0 +1,173 @@
+from comfy_execution.graph_utils import GraphBuilder, is_link
+from comfy_execution.graph import ExecutionBlocker
+from .tools import VariantSupport
+
+NUM_FLOW_SOCKETS = 5
+@VariantSupport()
+class TestWhileLoopOpen:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        inputs = {
+            "required": {
+                "condition": ("BOOLEAN", {"default": True}),
+            },
+            "optional": {
+            },
+        }
+        for i in range(NUM_FLOW_SOCKETS):
+            inputs["optional"][f"initial_value{i}"] = ("*",)
+        return inputs
+
+    RETURN_TYPES = tuple(["FLOW_CONTROL"] + ["*"] * NUM_FLOW_SOCKETS)
+    RETURN_NAMES = tuple(["FLOW_CONTROL"] + [f"value{i}" for i in range(NUM_FLOW_SOCKETS)])
+    FUNCTION = "while_loop_open"
+
+    CATEGORY = "Testing/Flow"
+
+    def while_loop_open(self, condition, **kwargs):
+        values = []
+        for i in range(NUM_FLOW_SOCKETS):
+            values.append(kwargs.get(f"initial_value{i}", None))
+        return tuple(["stub"] + values)
+
+@VariantSupport()
+class TestWhileLoopClose:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        inputs = {
+            "required": {
+                "flow_control": ("FLOW_CONTROL", {"rawLink": True}),
+                "condition": ("BOOLEAN", {"forceInput": True}),
+            },
+            "optional": {
+            },
+            "hidden": {
+                "dynprompt": "DYNPROMPT",
+                "unique_id": "UNIQUE_ID",
+            }
+        }
+        for i in range(NUM_FLOW_SOCKETS):
+            inputs["optional"][f"initial_value{i}"] = ("*",)
+        return inputs
+
+    RETURN_TYPES = tuple(["*"] * NUM_FLOW_SOCKETS)
+    RETURN_NAMES = tuple([f"value{i}" for i in range(NUM_FLOW_SOCKETS)])
+    FUNCTION = "while_loop_close"
+
+    CATEGORY = "Testing/Flow"
+
+    def explore_dependencies(self, node_id, dynprompt, upstream):
+        node_info = dynprompt.get_node(node_id)
+        if "inputs" not in node_info:
+            return
+        for k, v in node_info["inputs"].items():
+            if is_link(v):
+                parent_id = v[0]
+                if parent_id not in upstream:
+                    upstream[parent_id] = []
+                    self.explore_dependencies(parent_id, dynprompt, upstream)
+                upstream[parent_id].append(node_id)
+
+    def collect_contained(self, node_id, upstream, contained):
+        if node_id not in upstream:
+            return
+        for child_id in upstream[node_id]:
+            if child_id not in contained:
+                contained[child_id] = True
+                self.collect_contained(child_id, upstream, contained)
+
+
+    def while_loop_close(self, flow_control, condition, dynprompt=None, unique_id=None, **kwargs):
+        assert dynprompt is not None
+        if not condition:
+            # We're done with the loop
+            values = []
+            for i in range(NUM_FLOW_SOCKETS):
+                values.append(kwargs.get(f"initial_value{i}", None))
+            return tuple(values)
+
+        # We want to loop
+        upstream = {}
+        # Get the list of all nodes between the open and close nodes
+        self.explore_dependencies(unique_id, dynprompt, upstream)
+
+        contained = {}
+        open_node = flow_control[0]
+        self.collect_contained(open_node, upstream, contained)
+        contained[unique_id] = True
+        contained[open_node] = True
+
+        # We'll use the default prefix, but to avoid having node names grow exponentially in size,
+        # we'll use "Recurse" for the name of the recursively-generated copy of this node.
+        graph = GraphBuilder()
+        for node_id in contained:
+            original_node = dynprompt.get_node(node_id)
+            node = graph.node(original_node["class_type"], "Recurse" if node_id == unique_id else node_id)
+            node.set_override_display_id(node_id)
+        for node_id in contained:
+            original_node = dynprompt.get_node(node_id)
+            node = graph.lookup_node("Recurse" if node_id == unique_id else node_id)
+            assert node is not None
+            for k, v in original_node["inputs"].items():
+                if is_link(v) and v[0] in contained:
+                    parent = graph.lookup_node(v[0])
+                    assert parent is not None
+                    node.set_input(k, parent.out(v[1]))
+                else:
+                    node.set_input(k, v)
+        new_open = graph.lookup_node(open_node)
+        assert new_open is not None
+        for i in range(NUM_FLOW_SOCKETS):
+            key = f"initial_value{i}"
+            new_open.set_input(key, kwargs.get(key, None))
+        my_clone = graph.lookup_node("Recurse")
+        assert my_clone is not None
+        result = map(lambda x: my_clone.out(x), range(NUM_FLOW_SOCKETS))
+        return {
+            "result": tuple(result),
+            "expand": graph.finalize(),
+        }
+
+@VariantSupport()
+class TestExecutionBlockerNode:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        inputs = {
+            "required": {
+                "input": ("*",),
+                "block": ("BOOLEAN",),
+                "verbose": ("BOOLEAN", {"default": False}),
+            },
+        }
+        return inputs
+
+    RETURN_TYPES = ("*",)
+    RETURN_NAMES = ("output",)
+    FUNCTION = "execution_blocker"
+
+    CATEGORY = "Testing/Flow"
+
+    def execution_blocker(self, input, block, verbose):
+        if block:
+            return (ExecutionBlocker("Blocked Execution" if verbose else None),)
+        return (input,)
+
+FLOW_CONTROL_NODE_CLASS_MAPPINGS = {
+    "TestWhileLoopOpen": TestWhileLoopOpen,
+    "TestWhileLoopClose": TestWhileLoopClose,
+    "TestExecutionBlocker": TestExecutionBlockerNode,
+}
+FLOW_CONTROL_NODE_DISPLAY_NAME_MAPPINGS = {
+    "TestWhileLoopOpen": "While Loop Open",
+    "TestWhileLoopClose": "While Loop Close",
+    "TestExecutionBlocker": "Execution Blocker",
+}

tests/inference/testing_nodes/testing-pack/specific_tests.py

@@ -0,0 +1,362 @@
+import torch
+from .tools import VariantSupport
+from comfy_execution.graph_utils import GraphBuilder
+
+class TestLazyMixImages:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "image1": ("IMAGE",{"lazy": True}),
+                "image2": ("IMAGE",{"lazy": True}),
+                "mask": ("MASK",),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "mix"
+
+    CATEGORY = "Testing/Nodes"
+
+    def check_lazy_status(self, mask, image1, image2):
+        mask_min = mask.min()
+        mask_max = mask.max()
+        needed = []
+        if image1 is None and (mask_min != 1.0 or mask_max != 1.0):
+            needed.append("image1")
+        if image2 is None and (mask_min != 0.0 or mask_max != 0.0):
+            needed.append("image2")
+        return needed
+
+    # Not trying to handle different batch sizes here just to keep the demo simple
+    def mix(self, mask, image1, image2):
+        mask_min = mask.min()
+        mask_max = mask.max()
+        if mask_min == 0.0 and mask_max == 0.0:
+            return (image1,)
+        elif mask_min == 1.0 and mask_max == 1.0:
+            return (image2,)
+
+        if len(mask.shape) == 2:
+            mask = mask.unsqueeze(0)
+        if len(mask.shape) == 3:
+            mask = mask.unsqueeze(3)
+        if mask.shape[3] < image1.shape[3]:
+            mask = mask.repeat(1, 1, 1, image1.shape[3])
+
+        result = image1 * (1. - mask) + image2 * mask,
+        return (result[0],)
+
+class TestVariadicAverage:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "input1": ("IMAGE",),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "variadic_average"
+
+    CATEGORY = "Testing/Nodes"
+
+    def variadic_average(self, input1, **kwargs):
+        inputs = [input1]
+        while 'input' + str(len(inputs) + 1) in kwargs:
+            inputs.append(kwargs['input' + str(len(inputs) + 1)])
+        return (torch.stack(inputs).mean(dim=0),)
+
+
+class TestCustomIsChanged:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "image": ("IMAGE",),
+            },
+            "optional": {
+                "should_change": ("BOOL", {"default": False}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "custom_is_changed"
+
+    CATEGORY = "Testing/Nodes"
+
+    def custom_is_changed(self, image, should_change=False):
+        return (image,)
+    
+    @classmethod
+    def IS_CHANGED(cls, should_change=False, *args, **kwargs):
+        if should_change:
+            return float("NaN")
+        else:
+            return False
+
+class TestIsChangedWithConstants:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "image": ("IMAGE",),
+                "value": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "custom_is_changed"
+
+    CATEGORY = "Testing/Nodes"
+
+    def custom_is_changed(self, image, value):
+        return (image * value,)
+    
+    @classmethod
+    def IS_CHANGED(cls, image, value):
+        if image is None:
+            return value
+        else:
+            return image.mean().item() * value
+
+class TestCustomValidation1:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "input1": ("IMAGE,FLOAT",),
+                "input2": ("IMAGE,FLOAT",),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "custom_validation1"
+
+    CATEGORY = "Testing/Nodes"
+
+    def custom_validation1(self, input1, input2):
+        if isinstance(input1, float) and isinstance(input2, float):
+            result = torch.ones([1, 512, 512, 3]) * input1 * input2
+        else:
+            result = input1 * input2
+        return (result,)
+
+    @classmethod
+    def VALIDATE_INPUTS(cls, input1=None, input2=None):
+        if input1 is not None:
+            if not isinstance(input1, (torch.Tensor, float)):
+                return f"Invalid type of input1: {type(input1)}"
+        if input2 is not None:
+            if not isinstance(input2, (torch.Tensor, float)):
+                return f"Invalid type of input2: {type(input2)}"
+
+        return True
+
+class TestCustomValidation2:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "input1": ("IMAGE,FLOAT",),
+                "input2": ("IMAGE,FLOAT",),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "custom_validation2"
+
+    CATEGORY = "Testing/Nodes"
+
+    def custom_validation2(self, input1, input2):
+        if isinstance(input1, float) and isinstance(input2, float):
+            result = torch.ones([1, 512, 512, 3]) * input1 * input2
+        else:
+            result = input1 * input2
+        return (result,)
+
+    @classmethod
+    def VALIDATE_INPUTS(cls, input_types, input1=None, input2=None):
+        if input1 is not None:
+            if not isinstance(input1, (torch.Tensor, float)):
+                return f"Invalid type of input1: {type(input1)}"
+        if input2 is not None:
+            if not isinstance(input2, (torch.Tensor, float)):
+                return f"Invalid type of input2: {type(input2)}"
+
+        if 'input1' in input_types:
+            if input_types['input1'] not in ["IMAGE", "FLOAT"]:
+                return f"Invalid type of input1: {input_types['input1']}"
+        if 'input2' in input_types:
+            if input_types['input2'] not in ["IMAGE", "FLOAT"]:
+                return f"Invalid type of input2: {input_types['input2']}"
+
+        return True
+
+@VariantSupport()
+class TestCustomValidation3:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "input1": ("IMAGE,FLOAT",),
+                "input2": ("IMAGE,FLOAT",),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "custom_validation3"
+
+    CATEGORY = "Testing/Nodes"
+
+    def custom_validation3(self, input1, input2):
+        if isinstance(input1, float) and isinstance(input2, float):
+            result = torch.ones([1, 512, 512, 3]) * input1 * input2
+        else:
+            result = input1 * input2
+        return (result,)
+
+class TestCustomValidation4:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "input1": ("FLOAT",),
+                "input2": ("FLOAT",),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "custom_validation4"
+
+    CATEGORY = "Testing/Nodes"
+
+    def custom_validation4(self, input1, input2):
+        result = torch.ones([1, 512, 512, 3]) * input1 * input2
+        return (result,)
+
+    @classmethod
+    def VALIDATE_INPUTS(cls, input1, input2):
+        if input1 is not None:
+            if not isinstance(input1, float):
+                return f"Invalid type of input1: {type(input1)}"
+        if input2 is not None:
+            if not isinstance(input2, float):
+                return f"Invalid type of input2: {type(input2)}"
+
+        return True
+
+class TestCustomValidation5:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "input1": ("FLOAT", {"min": 0.0, "max": 1.0}),
+                "input2": ("FLOAT", {"min": 0.0, "max": 1.0}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "custom_validation5"
+
+    CATEGORY = "Testing/Nodes"
+
+    def custom_validation5(self, input1, input2):
+        value = input1 * input2
+        return (torch.ones([1, 512, 512, 3]) * value,)
+
+    @classmethod
+    def VALIDATE_INPUTS(cls, **kwargs):
+        if kwargs['input2'] == 7.0:
+            return "7s are not allowed. I've never liked 7s."
+        return True
+
+class TestDynamicDependencyCycle:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "input1": ("IMAGE",),
+                "input2": ("IMAGE",),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "dynamic_dependency_cycle"
+
+    CATEGORY = "Testing/Nodes"
+
+    def dynamic_dependency_cycle(self, input1, input2):
+        g = GraphBuilder()
+        mask = g.node("StubMask", value=0.5, height=512, width=512, batch_size=1)
+        mix1 = g.node("TestLazyMixImages", image1=input1, mask=mask.out(0))
+        mix2 = g.node("TestLazyMixImages", image1=mix1.out(0), image2=input2, mask=mask.out(0))
+
+        # Create the cyle
+        mix1.set_input("image2", mix2.out(0))
+
+        return {
+            "result": (mix2.out(0),),
+            "expand": g.finalize(),
+        }
+
+class TestMixedExpansionReturns:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "input1": ("FLOAT",),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE","IMAGE")
+    FUNCTION = "mixed_expansion_returns"
+
+    CATEGORY = "Testing/Nodes"
+
+    def mixed_expansion_returns(self, input1):
+        white_image = torch.ones([1, 512, 512, 3])
+        if input1 <= 0.1:
+            return (torch.ones([1, 512, 512, 3]) * 0.1, white_image)
+        elif input1 <= 0.2:
+            return {
+                "result": (torch.ones([1, 512, 512, 3]) * 0.2, white_image),
+            }
+        else:
+            g = GraphBuilder()
+            mask = g.node("StubMask", value=0.3, height=512, width=512, batch_size=1)
+            black = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+            white = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+            mix = g.node("TestLazyMixImages", image1=black.out(0), image2=white.out(0), mask=mask.out(0))
+            return {
+                "result": (mix.out(0), white_image),
+                "expand": g.finalize(),
+            }
+
+TEST_NODE_CLASS_MAPPINGS = {
+    "TestLazyMixImages": TestLazyMixImages,
+    "TestVariadicAverage": TestVariadicAverage,
+    "TestCustomIsChanged": TestCustomIsChanged,
+    "TestIsChangedWithConstants": TestIsChangedWithConstants,
+    "TestCustomValidation1": TestCustomValidation1,
+    "TestCustomValidation2": TestCustomValidation2,
+    "TestCustomValidation3": TestCustomValidation3,
+    "TestCustomValidation4": TestCustomValidation4,
+    "TestCustomValidation5": TestCustomValidation5,
+    "TestDynamicDependencyCycle": TestDynamicDependencyCycle,
+    "TestMixedExpansionReturns": TestMixedExpansionReturns,
+}
+
+TEST_NODE_DISPLAY_NAME_MAPPINGS = {
+    "TestLazyMixImages": "Lazy Mix Images",
+    "TestVariadicAverage": "Variadic Average",
+    "TestCustomIsChanged": "Custom IsChanged",
+    "TestIsChangedWithConstants": "IsChanged With Constants",
+    "TestCustomValidation1": "Custom Validation 1",
+    "TestCustomValidation2": "Custom Validation 2",
+    "TestCustomValidation3": "Custom Validation 3",
+    "TestCustomValidation4": "Custom Validation 4",
+    "TestCustomValidation5": "Custom Validation 5",
+    "TestDynamicDependencyCycle": "Dynamic Dependency Cycle",
+    "TestMixedExpansionReturns": "Mixed Expansion Returns",
+}

tests/inference/testing_nodes/testing-pack/stubs.py

@@ -0,0 +1,129 @@
+import torch
+
+class StubImage:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "content": (['WHITE', 'BLACK', 'NOISE'],),
+                "height": ("INT", {"default": 512, "min": 1, "max": 1024 ** 3, "step": 1}),
+                "width": ("INT", {"default": 512, "min": 1, "max": 4096 ** 3, "step": 1}),
+                "batch_size": ("INT", {"default": 1, "min": 1, "max": 1024 ** 3, "step": 1}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "stub_image"
+
+    CATEGORY = "Testing/Stub Nodes"
+
+    def stub_image(self, content, height, width, batch_size):
+        if content == "WHITE":
+            return (torch.ones(batch_size, height, width, 3),)
+        elif content == "BLACK":
+            return (torch.zeros(batch_size, height, width, 3),)
+        elif content == "NOISE":
+            return (torch.rand(batch_size, height, width, 3),)
+
+class StubConstantImage:
+    def __init__(self):
+        pass
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "height": ("INT", {"default": 512, "min": 1, "max": 1024 ** 3, "step": 1}),
+                "width": ("INT", {"default": 512, "min": 1, "max": 4096 ** 3, "step": 1}),
+                "batch_size": ("INT", {"default": 1, "min": 1, "max": 1024 ** 3, "step": 1}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "stub_constant_image"
+
+    CATEGORY = "Testing/Stub Nodes"
+
+    def stub_constant_image(self, value, height, width, batch_size):
+        return (torch.ones(batch_size, height, width, 3) * value,)
+
+class StubMask:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "height": ("INT", {"default": 512, "min": 1, "max": 1024 ** 3, "step": 1}),
+                "width": ("INT", {"default": 512, "min": 1, "max": 4096 ** 3, "step": 1}),
+                "batch_size": ("INT", {"default": 1, "min": 1, "max": 1024 ** 3, "step": 1}),
+            },
+        }
+
+    RETURN_TYPES = ("MASK",)
+    FUNCTION = "stub_mask"
+
+    CATEGORY = "Testing/Stub Nodes"
+
+    def stub_mask(self, value, height, width, batch_size):
+        return (torch.ones(batch_size, height, width) * value,)
+
+class StubInt:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": ("INT", {"default": 0, "min": -0xffffffff, "max": 0xffffffff, "step": 1}),
+            },
+        }
+
+    RETURN_TYPES = ("INT",)
+    FUNCTION = "stub_int"
+
+    CATEGORY = "Testing/Stub Nodes"
+
+    def stub_int(self, value):
+        return (value,)
+
+class StubFloat:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": ("FLOAT", {"default": 0.0, "min": -1.0e38, "max": 1.0e38, "step": 0.01}),
+            },
+        }
+
+    RETURN_TYPES = ("FLOAT",)
+    FUNCTION = "stub_float"
+
+    CATEGORY = "Testing/Stub Nodes"
+
+    def stub_float(self, value):
+        return (value,)
+
+TEST_STUB_NODE_CLASS_MAPPINGS = {
+    "StubImage": StubImage,
+    "StubConstantImage": StubConstantImage,
+    "StubMask": StubMask,
+    "StubInt": StubInt,
+    "StubFloat": StubFloat,
+}
+TEST_STUB_NODE_DISPLAY_NAME_MAPPINGS = {
+    "StubImage": "Stub Image",
+    "StubConstantImage": "Stub Constant Image",
+    "StubMask": "Stub Mask",
+    "StubInt": "Stub Int",
+    "StubFloat": "Stub Float",
+}

tests/inference/testing_nodes/testing-pack/tools.py

@@ -0,0 +1,53 @@
+
+def MakeSmartType(t):
+    if isinstance(t, str):
+        return SmartType(t)
+    return t
+
+class SmartType(str):
+    def __ne__(self, other):
+        if self == "*" or other == "*":
+            return False
+        selfset = set(self.split(','))
+        otherset = set(other.split(','))
+        return not selfset.issubset(otherset)
+
+def VariantSupport():
+    def decorator(cls):
+        if hasattr(cls, "INPUT_TYPES"):
+            old_input_types = getattr(cls, "INPUT_TYPES")
+            def new_input_types(*args, **kwargs):
+                types = old_input_types(*args, **kwargs)
+                for category in ["required", "optional"]:
+                    if category not in types:
+                        continue
+                    for key, value in types[category].items():
+                        if isinstance(value, tuple):
+                            types[category][key] = (MakeSmartType(value[0]),) + value[1:]
+                return types
+            setattr(cls, "INPUT_TYPES", new_input_types)
+        if hasattr(cls, "RETURN_TYPES"):
+            old_return_types = cls.RETURN_TYPES
+            setattr(cls, "RETURN_TYPES", tuple(MakeSmartType(x) for x in old_return_types))
+        if hasattr(cls, "VALIDATE_INPUTS"):
+            # Reflection is used to determine what the function signature is, so we can't just change the function signature
+            raise NotImplementedError("VariantSupport does not support VALIDATE_INPUTS yet")
+        else:
+            def validate_inputs(input_types):
+                inputs = cls.INPUT_TYPES()
+                for key, value in input_types.items():
+                    if isinstance(value, SmartType):
+                        continue
+                    if "required" in inputs and key in inputs["required"]:
+                        expected_type = inputs["required"][key][0]
+                    elif "optional" in inputs and key in inputs["optional"]:
+                        expected_type = inputs["optional"][key][0]
+                    else:
+                        expected_type = None
+                    if expected_type is not None and MakeSmartType(value) != expected_type:
+                        return f"Invalid type of {key}: {value} (expected {expected_type})"
+                return True
+            setattr(cls, "VALIDATE_INPUTS", validate_inputs)
+        return cls
+    return decorator
+

tests/inference/testing_nodes/testing-pack/util.py

@@ -0,0 +1,364 @@
+from comfy_execution.graph_utils import GraphBuilder
+from .tools import VariantSupport
+
+@VariantSupport()
+class TestAccumulateNode:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "to_add": ("*",),
+            },
+            "optional": {
+                "accumulation": ("ACCUMULATION",),
+            },
+        }
+
+    RETURN_TYPES = ("ACCUMULATION",)
+    FUNCTION = "accumulate"
+
+    CATEGORY = "Testing/Lists"
+
+    def accumulate(self, to_add, accumulation = None):
+        if accumulation is None:
+            value = [to_add]
+        else:
+            value = accumulation["accum"] + [to_add]
+        return ({"accum": value},)
+
+@VariantSupport()
+class TestAccumulationHeadNode:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "accumulation": ("ACCUMULATION",),
+            },
+        }
+
+    RETURN_TYPES = ("ACCUMULATION", "*",)
+    FUNCTION = "accumulation_head"
+
+    CATEGORY = "Testing/Lists"
+
+    def accumulation_head(self, accumulation):
+        accum = accumulation["accum"]
+        if len(accum) == 0:
+            return (accumulation, None)
+        else:
+            return ({"accum": accum[1:]}, accum[0])
+
+class TestAccumulationTailNode:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "accumulation": ("ACCUMULATION",),
+            },
+        }
+
+    RETURN_TYPES = ("ACCUMULATION", "*",)
+    FUNCTION = "accumulation_tail"
+
+    CATEGORY = "Testing/Lists"
+
+    def accumulation_tail(self, accumulation):
+        accum = accumulation["accum"]
+        if len(accum) == 0:
+            return (None, accumulation)
+        else:
+            return ({"accum": accum[:-1]}, accum[-1])
+
+@VariantSupport()
+class TestAccumulationToListNode:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "accumulation": ("ACCUMULATION",),
+            },
+        }
+
+    RETURN_TYPES = ("*",)
+    OUTPUT_IS_LIST = (True,)
+
+    FUNCTION = "accumulation_to_list"
+
+    CATEGORY = "Testing/Lists"
+
+    def accumulation_to_list(self, accumulation):
+        return (accumulation["accum"],)
+
+@VariantSupport()
+class TestListToAccumulationNode:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "list": ("*",),
+            },
+        }
+
+    RETURN_TYPES = ("ACCUMULATION",)
+    INPUT_IS_LIST = (True,)
+
+    FUNCTION = "list_to_accumulation"
+
+    CATEGORY = "Testing/Lists"
+
+    def list_to_accumulation(self, list):
+        return ({"accum": list},)
+
+@VariantSupport()
+class TestAccumulationGetLengthNode:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "accumulation": ("ACCUMULATION",),
+            },
+        }
+
+    RETURN_TYPES = ("INT",)
+
+    FUNCTION = "accumlength"
+
+    CATEGORY = "Testing/Lists"
+
+    def accumlength(self, accumulation):
+        return (len(accumulation['accum']),)
+        
+@VariantSupport()
+class TestAccumulationGetItemNode:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "accumulation": ("ACCUMULATION",),
+                "index": ("INT", {"default":0, "step":1})
+            },
+        }
+
+    RETURN_TYPES = ("*",)
+
+    FUNCTION = "get_item"
+
+    CATEGORY = "Testing/Lists"
+
+    def get_item(self, accumulation, index):
+        return (accumulation['accum'][index],)
+        
+@VariantSupport()
+class TestAccumulationSetItemNode:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "accumulation": ("ACCUMULATION",),
+                "index": ("INT", {"default":0, "step":1}),
+                "value": ("*",),
+            },
+        }
+
+    RETURN_TYPES = ("ACCUMULATION",)
+
+    FUNCTION = "set_item"
+
+    CATEGORY = "Testing/Lists"
+
+    def set_item(self, accumulation, index, value):
+        new_accum = accumulation['accum'][:]
+        new_accum[index] = value
+        return ({"accum": new_accum},)
+
+class TestIntMathOperation:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "a": ("INT", {"default": 0, "min": -0xffffffffffffffff, "max": 0xffffffffffffffff, "step": 1}),
+                "b": ("INT", {"default": 0, "min": -0xffffffffffffffff, "max": 0xffffffffffffffff, "step": 1}),
+                "operation": (["add", "subtract", "multiply", "divide", "modulo", "power"],),
+            },
+        }
+
+    RETURN_TYPES = ("INT",)
+    FUNCTION = "int_math_operation"
+
+    CATEGORY = "Testing/Logic"
+
+    def int_math_operation(self, a, b, operation):
+        if operation == "add":
+            return (a + b,)
+        elif operation == "subtract":
+            return (a - b,)
+        elif operation == "multiply":
+            return (a * b,)
+        elif operation == "divide":
+            return (a // b,)
+        elif operation == "modulo":
+            return (a % b,)
+        elif operation == "power":
+            return (a ** b,)
+
+
+from .flow_control import NUM_FLOW_SOCKETS
+@VariantSupport()
+class TestForLoopOpen:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "remaining": ("INT", {"default": 1, "min": 0, "max": 100000, "step": 1}),
+            },
+            "optional": {
+                f"initial_value{i}": ("*",) for i in range(1, NUM_FLOW_SOCKETS)
+            },
+            "hidden": {
+                "initial_value0": ("*",)
+            }
+        }
+
+    RETURN_TYPES = tuple(["FLOW_CONTROL", "INT",] + ["*"] * (NUM_FLOW_SOCKETS-1))
+    RETURN_NAMES = tuple(["flow_control", "remaining"] + [f"value{i}" for i in range(1, NUM_FLOW_SOCKETS)])
+    FUNCTION = "for_loop_open"
+
+    CATEGORY = "Testing/Flow"
+
+    def for_loop_open(self, remaining, **kwargs):
+        graph = GraphBuilder()
+        if "initial_value0" in kwargs:
+            remaining = kwargs["initial_value0"]
+        while_open = graph.node("TestWhileLoopOpen", condition=remaining, initial_value0=remaining, **{(f"initial_value{i}"): kwargs.get(f"initial_value{i}", None) for i in range(1, NUM_FLOW_SOCKETS)})
+        outputs = [kwargs.get(f"initial_value{i}", None) for i in range(1, NUM_FLOW_SOCKETS)]
+        return {
+            "result": tuple(["stub", remaining] + outputs),
+            "expand": graph.finalize(),
+        }
+
+@VariantSupport()
+class TestForLoopClose:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "flow_control": ("FLOW_CONTROL", {"rawLink": True}),
+            },
+            "optional": {
+                f"initial_value{i}": ("*",{"rawLink": True}) for i in range(1, NUM_FLOW_SOCKETS)
+            },
+        }
+
+    RETURN_TYPES = tuple(["*"] * (NUM_FLOW_SOCKETS-1))
+    RETURN_NAMES = tuple([f"value{i}" for i in range(1, NUM_FLOW_SOCKETS)])
+    FUNCTION = "for_loop_close"
+
+    CATEGORY = "Testing/Flow"
+
+    def for_loop_close(self, flow_control, **kwargs):
+        graph = GraphBuilder()
+        while_open = flow_control[0]
+        sub = graph.node("TestIntMathOperation", operation="subtract", a=[while_open,1], b=1)
+        cond = graph.node("TestToBoolNode", value=sub.out(0))
+        input_values = {f"initial_value{i}": kwargs.get(f"initial_value{i}", None) for i in range(1, NUM_FLOW_SOCKETS)}
+        while_close = graph.node("TestWhileLoopClose",
+                flow_control=flow_control,
+                condition=cond.out(0),
+                initial_value0=sub.out(0),
+                **input_values)
+        return {
+            "result": tuple([while_close.out(i) for i in range(1, NUM_FLOW_SOCKETS)]),
+            "expand": graph.finalize(),
+        }
+
+NUM_LIST_SOCKETS = 10
+@VariantSupport()
+class TestMakeListNode:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value1": ("*",),
+            },
+            "optional": {
+                f"value{i}": ("*",) for i in range(1, NUM_LIST_SOCKETS)
+            },
+        }
+
+    RETURN_TYPES = ("*",)
+    FUNCTION = "make_list"
+    OUTPUT_IS_LIST = (True,)
+
+    CATEGORY = "Testing/Lists"
+
+    def make_list(self, **kwargs):
+        result = []
+        for i in range(NUM_LIST_SOCKETS):
+            if f"value{i}" in kwargs:
+                result.append(kwargs[f"value{i}"])
+        return (result,)
+
+UTILITY_NODE_CLASS_MAPPINGS = {
+    "TestAccumulateNode": TestAccumulateNode,
+    "TestAccumulationHeadNode": TestAccumulationHeadNode,
+    "TestAccumulationTailNode": TestAccumulationTailNode,
+    "TestAccumulationToListNode": TestAccumulationToListNode,
+    "TestListToAccumulationNode": TestListToAccumulationNode,
+    "TestAccumulationGetLengthNode": TestAccumulationGetLengthNode,
+    "TestAccumulationGetItemNode": TestAccumulationGetItemNode,
+    "TestAccumulationSetItemNode": TestAccumulationSetItemNode,
+    "TestForLoopOpen": TestForLoopOpen,
+    "TestForLoopClose": TestForLoopClose,
+    "TestIntMathOperation": TestIntMathOperation,
+    "TestMakeListNode": TestMakeListNode,
+}
+UTILITY_NODE_DISPLAY_NAME_MAPPINGS = {
+    "TestAccumulateNode": "Accumulate",
+    "TestAccumulationHeadNode": "Accumulation Head",
+    "TestAccumulationTailNode": "Accumulation Tail",
+    "TestAccumulationToListNode": "Accumulation to List",
+    "TestListToAccumulationNode": "List to Accumulation",
+    "TestAccumulationGetLengthNode": "Accumulation Get Length",
+    "TestAccumulationGetItemNode": "Accumulation Get Item",
+    "TestAccumulationSetItemNode": "Accumulation Set Item",
+    "TestForLoopOpen": "For Loop Open",
+    "TestForLoopClose": "For Loop Close",
+    "TestIntMathOperation": "Int Math Operation",
+    "TestMakeListNode": "Make List",
+}

web/assets/index-DjWyclij.css

@@ -0,0 +1,149 @@
+.comfy-group-manage {
+	background: var(--bg-color);
+	color: var(--fg-color);
+	padding: 0;
+	font-family: Arial, Helvetica, sans-serif;
+	border-color: black;
+	margin: 20vh auto;
+	max-height: 60vh;
+}
+.comfy-group-manage-outer {
+	max-height: 60vh;
+	min-width: 500px;
+	display: flex;
+	flex-direction: column;
+}
+.comfy-group-manage-outer > header {
+	display: flex;
+	align-items: center;
+	gap: 10px;
+	justify-content: space-between;
+	background: var(--comfy-menu-bg);
+	padding: 15px 20px;
+}
+.comfy-group-manage-outer > header select {
+	background: var(--comfy-input-bg);
+	border: 1px solid var(--border-color);
+	color: var(--input-text);
+	padding: 5px 10px;
+	border-radius: 5px;
+}
+.comfy-group-manage h2 {
+	margin: 0;
+	font-weight: normal;
+}
+.comfy-group-manage main {
+	display: flex;
+	overflow: hidden;
+}
+.comfy-group-manage .drag-handle {
+	font-weight: bold;
+}
+.comfy-group-manage-list {
+	border-right: 1px solid var(--comfy-menu-bg);
+}
+.comfy-group-manage-list ul {
+	margin: 40px 0 0;
+	padding: 0;
+	list-style: none;
+}
+.comfy-group-manage-list-items {
+	max-height: calc(100% - 40px);
+	overflow-y: scroll;
+	overflow-x: hidden;
+}
+.comfy-group-manage-list li {
+	display: flex;
+	padding: 10px 20px 10px 10px;
+	cursor: pointer;
+	align-items: center;
+	gap: 5px;
+}
+.comfy-group-manage-list div {
+	display: flex;
+	flex-direction: column;
+}
+.comfy-group-manage-list li:not(.selected):hover div {
+	text-decoration: underline;
+}
+.comfy-group-manage-list li.selected {
+	background: var(--border-color);
+}
+.comfy-group-manage-list li span {
+	opacity: 0.7;
+	font-size: smaller;
+}
+.comfy-group-manage-node {
+	flex: auto;
+	background: var(--border-color);
+	display: flex;
+	flex-direction: column;
+}
+.comfy-group-manage-node > div {
+	overflow: auto;
+}
+.comfy-group-manage-node header {
+	display: flex;
+	background: var(--bg-color);
+	height: 40px;
+}
+.comfy-group-manage-node header a {
+	text-align: center;
+	flex: auto;
+	border-right: 1px solid var(--comfy-menu-bg);
+	border-bottom: 1px solid var(--comfy-menu-bg);
+	padding: 10px;
+	cursor: pointer;
+	font-size: 15px;
+}
+.comfy-group-manage-node header a:last-child {
+    border-right: none;
+}
+.comfy-group-manage-node header a:not(.active):hover {
+	text-decoration: underline;
+}
+.comfy-group-manage-node header a.active {
+	background: var(--border-color);
+	border-bottom: none;
+}
+.comfy-group-manage-node-page {
+	display: none;
+	overflow: auto;
+}
+.comfy-group-manage-node-page.active {
+	display: block;
+}
+.comfy-group-manage-node-page div {
+	padding: 10px;
+	display: flex;
+	align-items: center;
+	gap: 10px;
+}
+.comfy-group-manage-node-page input {
+	border: none;
+	color: var(--input-text);
+	background: var(--comfy-input-bg);
+	padding: 5px 10px;
+}
+.comfy-group-manage-node-page input[type="text"] {
+	flex: auto;
+}
+.comfy-group-manage-node-page label {
+	display: flex;
+	gap: 5px;
+	align-items: center;
+}
+.comfy-group-manage footer {
+	border-top: 1px solid var(--comfy-menu-bg);
+	padding: 10px;
+	display: flex;
+	gap: 10px;
+}
+.comfy-group-manage footer button {
+	font-size: 14px;
+	padding: 5px 10px;
+	border-radius: 0;
+}
+.comfy-group-manage footer button:first-child {
+	margin-right: auto;
+}

web/assets/userSelection-BGzn1LuN.css

@@ -0,0 +1,136 @@
+.comfy-user-selection {
+    width: 100vw;
+    height: 100vh;
+    position: absolute;
+    top: 0;
+    left: 0;
+    z-index: 999;
+    display: flex;
+    align-items: center;
+    justify-content: center;
+    font-family: sans-serif;
+    background: linear-gradient(var(--tr-even-bg-color), var(--tr-odd-bg-color));
+}
+
+.comfy-user-selection-inner {
+    background: var(--comfy-menu-bg);
+    margin-top: -30vh;
+    padding: 20px 40px;
+    border-radius: 10px;
+    min-width: 365px;
+    position: relative;
+    box-shadow: 0 0 20px rgba(0, 0, 0, 0.3);
+}
+
+.comfy-user-selection-inner form {
+    width: 100%;
+    display: flex;
+    flex-direction: column;
+    align-items: center;
+}
+
+.comfy-user-selection-inner h1 {
+    margin: 10px 0 30px 0;
+    font-weight: normal;
+}
+
+.comfy-user-selection-inner label {
+    display: flex;
+    flex-direction: column;
+    width: 100%;
+}
+
+.comfy-user-selection input,
+.comfy-user-selection select {
+    background-color: var(--comfy-input-bg);
+    color: var(--input-text);
+    border: 0;
+    border-radius: 5px;
+    padding: 5px;
+    margin-top: 10px;
+}
+
+.comfy-user-selection input::-moz-placeholder {
+    color: var(--descrip-text);
+    opacity: 1;
+}
+
+.comfy-user-selection input::placeholder {
+    color: var(--descrip-text);
+    opacity: 1;
+}
+
+.comfy-user-existing {
+    width: 100%;
+}
+
+.no-users .comfy-user-existing {
+    display: none;
+}
+
+.comfy-user-selection-inner .or-separator {
+    margin: 10px 0;
+    padding: 10px;
+    display: block;
+    width: 100%;
+    color: var(--descrip-text);
+    overflow: hidden;
+    text-align: center;
+    margin-left: -10px;
+}
+
+.comfy-user-selection-inner .or-separator::before,
+.comfy-user-selection-inner .or-separator::after {
+    content: "";
+    background-color: var(--border-color);
+    position: relative;
+    height: 1px;
+    vertical-align: middle;
+    display: inline-block;
+    width: calc(50% - 20px);
+    top: -1px;
+}
+
+.comfy-user-selection-inner .or-separator::before {
+    right: 10px;
+    margin-left: -50%;
+}
+
+.comfy-user-selection-inner .or-separator::after {
+    left: 10px;
+    margin-right: -50%;
+}
+
+.comfy-user-selection-inner section {
+    width: 100%;
+    padding: 10px;
+    margin: -10px;
+    transition: background-color 0.2s;
+}
+
+.comfy-user-selection-inner section.selected {
+    background: var(--border-color);
+    border-radius: 5px;
+}
+
+.comfy-user-selection-inner footer {
+    display: flex;
+    flex-direction: column;
+    align-items: center;
+    margin-top: 20px;
+}
+
+.comfy-user-selection-inner .comfy-user-error {
+    color: var(--error-text);
+    margin-bottom: 10px;
+}
+
+.comfy-user-button-next {
+    font-size: 16px;
+    padding: 6px 10px;
+    width: 100px;
+    display: flex;
+    gap: 5px;
+    align-items: center;
+    justify-content: center;
+}

web/assets/userSelection-GRU1gtOt.js

@@ -0,0 +1,142 @@
+var __defProp = Object.defineProperty;
+var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
+var __async = (__this, __arguments, generator) => {
+  return new Promise((resolve, reject) => {
+    var fulfilled = (value) => {
+      try {
+        step(generator.next(value));
+      } catch (e) {
+        reject(e);
+      }
+    };
+    var rejected = (value) => {
+      try {
+        step(generator.throw(value));
+      } catch (e) {
+        reject(e);
+      }
+    };
+    var step = (x) => x.done ? resolve(x.value) : Promise.resolve(x.value).then(fulfilled, rejected);
+    step((generator = generator.apply(__this, __arguments)).next());
+  });
+};
+import { j as createSpinner, g as api, $ as $el } from "./index-CaD4RONs.js";
+const _UserSelectionScreen = class _UserSelectionScreen {
+  show(users, user) {
+    return __async(this, null, function* () {
+      const userSelection = document.getElementById("comfy-user-selection");
+      userSelection.style.display = "";
+      return new Promise((resolve) => {
+        const input = userSelection.getElementsByTagName("input")[0];
+        const select = userSelection.getElementsByTagName("select")[0];
+        const inputSection = input.closest("section");
+        const selectSection = select.closest("section");
+        const form = userSelection.getElementsByTagName("form")[0];
+        const error = userSelection.getElementsByClassName("comfy-user-error")[0];
+        const button = userSelection.getElementsByClassName(
+          "comfy-user-button-next"
+        )[0];
+        let inputActive = null;
+        input.addEventListener("focus", () => {
+          inputSection.classList.add("selected");
+          selectSection.classList.remove("selected");
+          inputActive = true;
+        });
+        select.addEventListener("focus", () => {
+          inputSection.classList.remove("selected");
+          selectSection.classList.add("selected");
+          inputActive = false;
+          select.style.color = "";
+        });
+        select.addEventListener("blur", () => {
+          if (!select.value) {
+            select.style.color = "var(--descrip-text)";
+          }
+        });
+        form.addEventListener("submit", (e) => __async(this, null, function* () {
+          var _a, _b, _c;
+          e.preventDefault();
+          if (inputActive == null) {
+            error.textContent = "Please enter a username or select an existing user.";
+          } else if (inputActive) {
+            const username = input.value.trim();
+            if (!username) {
+              error.textContent = "Please enter a username.";
+              return;
+            }
+            input.disabled = select.disabled = // @ts-expect-error
+            input.readonly = // @ts-expect-error
+            select.readonly = true;
+            const spinner = createSpinner();
+            button.prepend(spinner);
+            try {
+              const resp = yield api.createUser(username);
+              if (resp.status >= 300) {
+                let message = "Error creating user: " + resp.status + " " + resp.statusText;
+                try {
+                  const res = yield resp.json();
+                  if (res.error) {
+                    message = res.error;
+                  }
+                } catch (error2) {
+                }
+                throw new Error(message);
+              }
+              resolve({ username, userId: yield resp.json(), created: true });
+            } catch (err) {
+              spinner.remove();
+              error.textContent = (_c = (_b = (_a = err.message) != null ? _a : err.statusText) != null ? _b : err) != null ? _c : "An unknown error occurred.";
+              input.disabled = select.disabled = // @ts-expect-error
+              input.readonly = // @ts-expect-error
+              select.readonly = false;
+              return;
+            }
+          } else if (!select.value) {
+            error.textContent = "Please select an existing user.";
+            return;
+          } else {
+            resolve({
+              username: users[select.value],
+              userId: select.value,
+              created: false
+            });
+          }
+        }));
+        if (user) {
+          const name = localStorage["Comfy.userName"];
+          if (name) {
+            input.value = name;
+          }
+        }
+        if (input.value) {
+          input.focus();
+        }
+        const userIds = Object.keys(users != null ? users : {});
+        if (userIds.length) {
+          for (const u of userIds) {
+            $el("option", { textContent: users[u], value: u, parent: select });
+          }
+          select.style.color = "var(--descrip-text)";
+          if (select.value) {
+            select.focus();
+          }
+        } else {
+          userSelection.classList.add("no-users");
+          input.focus();
+        }
+      }).then((r) => {
+        userSelection.remove();
+        return r;
+      });
+    });
+  }
+};
+__name(_UserSelectionScreen, "UserSelectionScreen");
+let UserSelectionScreen = _UserSelectionScreen;
+window.comfyAPI = window.comfyAPI || {};
+window.comfyAPI.userSelection = window.comfyAPI.userSelection || {};
+window.comfyAPI.userSelection.UserSelectionScreen = UserSelectionScreen;
+export {
+  UserSelectionScreen
+};
+//# sourceMappingURL=userSelection-GRU1gtOt.js.map

web/extensions/core/clipspace.js

@@ -1,166 +1,2 @@
-import { app } from "../../scripts/app.js";
+// Shim for extensions\core\clipspace.ts
-import { ComfyDialog, $el } from "../../scripts/ui.js";
+export const ClipspaceDialog = window.comfyAPI.clipspace.ClipspaceDialog;
-import { ComfyApp } from "../../scripts/app.js";
-
-export class ClipspaceDialog extends ComfyDialog {
-	static items = [];
-	static instance = null;
-
-	static registerButton(name, contextPredicate, callback) {
-		const item =
-			$el("button", {
-				type: "button",
-				textContent: name,
-				contextPredicate: contextPredicate,
-				onclick: callback
-			})
-
-		ClipspaceDialog.items.push(item);
-	}
-
-	static invalidatePreview() {
-		if(ComfyApp.clipspace && ComfyApp.clipspace.imgs && ComfyApp.clipspace.imgs.length > 0) {
-			const img_preview = document.getElementById("clipspace_preview");
-			if(img_preview) {
-				img_preview.src = ComfyApp.clipspace.imgs[ComfyApp.clipspace['selectedIndex']].src;
-				img_preview.style.maxHeight = "100%";
-				img_preview.style.maxWidth = "100%";
-			}
-		}
-	}
-
-	static invalidate() {
-		if(ClipspaceDialog.instance) {
-			const self = ClipspaceDialog.instance;
-			// allow reconstruct controls when copying from non-image to image content.
-			const children = $el("div.comfy-modal-content", [ self.createImgSettings(), ...self.createButtons() ]);
-
-			if(self.element) {
-				// update
-				self.element.removeChild(self.element.firstChild);
-				self.element.appendChild(children);
-			}
-			else {
-				// new
-				self.element = $el("div.comfy-modal", { parent: document.body }, [children,]);
-			}
-
-			if(self.element.children[0].children.length <= 1) {
-				self.element.children[0].appendChild($el("p", {}, ["Unable to find the features to edit content of a format stored in the current Clipspace."]));
-			}
-
-			ClipspaceDialog.invalidatePreview();
-		}
-	}
-
-	constructor() {
-		super();
-	}
-
-	createButtons(self) {
-		const buttons = [];
-
-		for(let idx in ClipspaceDialog.items) {
-			const item = ClipspaceDialog.items[idx];
-			if(!item.contextPredicate || item.contextPredicate())
-				buttons.push(ClipspaceDialog.items[idx]);
-		}
-
-		buttons.push(
-			$el("button", {
-				type: "button",
-				textContent: "Close",
-				onclick: () => { this.close(); }
-			})
-		);
-
-		return buttons;
-	}
-
-	createImgSettings() {
-		if(ComfyApp.clipspace.imgs) {
-			const combo_items = [];
-			const imgs = ComfyApp.clipspace.imgs;
-
-			for(let i=0; i < imgs.length; i++) {
-				combo_items.push($el("option", {value:i}, [`${i}`]));
-			}
-
-			const combo1 = $el("select",
-				{id:"clipspace_img_selector", onchange:(event) => {
-					ComfyApp.clipspace['selectedIndex'] = event.target.selectedIndex;
-					ClipspaceDialog.invalidatePreview();
-				} }, combo_items);
-
-			const row1 =
-				$el("tr", {},
-						[
-							$el("td", {}, [$el("font", {color:"white"}, ["Select Image"])]),
-							$el("td", {}, [combo1])
-						]);
-
-
-			const combo2 = $el("select",
-								{id:"clipspace_img_paste_mode", onchange:(event) => {
-									ComfyApp.clipspace['img_paste_mode'] = event.target.value;
-								} },
-								[
-									$el("option", {value:'selected'}, 'selected'),
-									$el("option", {value:'all'}, 'all')
-								]);
-			combo2.value = ComfyApp.clipspace['img_paste_mode'];
-
-			const row2 =
-				$el("tr", {},
-						[
-							$el("td", {}, [$el("font", {color:"white"}, ["Paste Mode"])]),
-							$el("td", {}, [combo2])
-						]);
-
-			const td = $el("td", {align:'center', width:'100px', height:'100px', colSpan:'2'},
-								[ $el("img",{id:"clipspace_preview", ondragstart:() => false},[]) ]);
-
-			const row3 =
-				$el("tr", {}, [td]);
-
-			return $el("table", {}, [row1, row2, row3]);
-		}
-		else {
-			return [];
-		}
-	}
-
-	createImgPreview() {
-		if(ComfyApp.clipspace.imgs) {
-			return $el("img",{id:"clipspace_preview", ondragstart:() => false});
-		}
-		else
-			return [];
-	}
-
-	show() {
-		const img_preview = document.getElementById("clipspace_preview");
-		ClipspaceDialog.invalidate();
-		
-		this.element.style.display = "block";
-	}
-}
-
-app.registerExtension({
-	name: "Comfy.Clipspace",
-	init(app) {
-		app.openClipspace =
-			function () {
-				if(!ClipspaceDialog.instance) {
-					ClipspaceDialog.instance = new ClipspaceDialog(app);
-					ComfyApp.clipspace_invalidate_handler = ClipspaceDialog.invalidate;
-				}
-
-				if(ComfyApp.clipspace) {
-					ClipspaceDialog.instance.show();
-				}
-				else
-					app.ui.dialog.show("Clipspace is Empty!");
-			};
-	}
-});

web/extensions/core/groupNodeManage.js

@@ -1,422 +1,2 @@
-import { $el, ComfyDialog } from "../../scripts/ui.js";
+// Shim for extensions\core\groupNodeManage.ts
-import { DraggableList } from "../../scripts/ui/draggableList.js";
+export const ManageGroupDialog = window.comfyAPI.groupNodeManage.ManageGroupDialog;
-import { addStylesheet } from "../../scripts/utils.js";
-import { GroupNodeConfig, GroupNodeHandler } from "./groupNode.js";
-
-addStylesheet(import.meta.url);
-
-const ORDER = Symbol();
-
-function merge(target, source) {
-	if (typeof target === "object" && typeof source === "object") {
-		for (const key in source) {
-			const sv = source[key];
-			if (typeof sv === "object") {
-				let tv = target[key];
-				if (!tv) tv = target[key] = {};
-				merge(tv, source[key]);
-			} else {
-				target[key] = sv;
-			}
-		}
-	}
-
-	return target;
-}
-
-export class ManageGroupDialog extends ComfyDialog {
-	/** @type { Record<"Inputs" | "Outputs" | "Widgets", {tab: HTMLAnchorElement, page: HTMLElement}> } */
-	tabs = {};
-	/** @type { number | null | undefined } */
-	selectedNodeIndex;
-	/** @type { keyof ManageGroupDialog["tabs"] } */
-	selectedTab = "Inputs";
-	/** @type { string | undefined } */
-	selectedGroup;
-
-	/** @type { Record<string, Record<string, Record<string, { name?: string | undefined, visible?: boolean | undefined }>>> } */
-	modifications = {};
-
-	get selectedNodeInnerIndex() {
-		return +this.nodeItems[this.selectedNodeIndex].dataset.nodeindex;
-	}
-
-	constructor(app) {
-		super();
-		this.app = app;
-		this.element = $el("dialog.comfy-group-manage", {
-			parent: document.body,
-		});
-	}
-
-	changeTab(tab) {
-		this.tabs[this.selectedTab].tab.classList.remove("active");
-		this.tabs[this.selectedTab].page.classList.remove("active");
-		this.tabs[tab].tab.classList.add("active");
-		this.tabs[tab].page.classList.add("active");
-		this.selectedTab = tab;
-	}
-
-	changeNode(index, force) {
-		if (!force && this.selectedNodeIndex === index) return;
-
-		if (this.selectedNodeIndex != null) {
-			this.nodeItems[this.selectedNodeIndex].classList.remove("selected");
-		}
-		this.nodeItems[index].classList.add("selected");
-		this.selectedNodeIndex = index;
-
-		if (!this.buildInputsPage() && this.selectedTab === "Inputs") {
-			this.changeTab("Widgets");
-		}
-		if (!this.buildWidgetsPage() && this.selectedTab === "Widgets") {
-			this.changeTab("Outputs");
-		}
-		if (!this.buildOutputsPage() && this.selectedTab === "Outputs") {
-			this.changeTab("Inputs");
-		}
-
-		this.changeTab(this.selectedTab);
-	}
-
-	getGroupData() {
-		this.groupNodeType = LiteGraph.registered_node_types["workflow/" + this.selectedGroup];
-		this.groupNodeDef = this.groupNodeType.nodeData;
-		this.groupData = GroupNodeHandler.getGroupData(this.groupNodeType);
-	}
-
-	changeGroup(group, reset = true) {
-		this.selectedGroup = group;
-		this.getGroupData();
-
-		const nodes = this.groupData.nodeData.nodes;
-		this.nodeItems = nodes.map((n, i) =>
-			$el(
-				"li.draggable-item",
-				{
-					dataset: {
-						nodeindex: n.index + "",
-					},
-					onclick: () => {
-						this.changeNode(i);
-					},
-				},
-				[
-					$el("span.drag-handle"),
-					$el(
-						"div",
-						{
-							textContent: n.title ?? n.type,
-						},
-						n.title
-							? $el("span", {
-									textContent: n.type,
-							  })
-							: []
-					),
-				]
-			)
-		);
-
-		this.innerNodesList.replaceChildren(...this.nodeItems);
-
-		if (reset) {
-			this.selectedNodeIndex = null;
-			this.changeNode(0);
-		} else {
-			const items = this.draggable.getAllItems();
-			let index = items.findIndex(item => item.classList.contains("selected"));
-			if(index === -1) index = this.selectedNodeIndex;
-			this.changeNode(index, true);
-		}
-
-		const ordered = [...nodes];
-		this.draggable?.dispose();
-		this.draggable = new DraggableList(this.innerNodesList, "li");
-		this.draggable.addEventListener("dragend", ({ detail: { oldPosition, newPosition } }) => {
-			if (oldPosition === newPosition) return;
-			ordered.splice(newPosition, 0, ordered.splice(oldPosition, 1)[0]);
-			for (let i = 0; i < ordered.length; i++) {
-				this.storeModification({ nodeIndex: ordered[i].index, section: ORDER, prop: "order", value: i });
-			}
-		});
-	}
-
-	storeModification({ nodeIndex, section, prop, value }) {
-		const groupMod = (this.modifications[this.selectedGroup] ??= {});
-		const nodesMod = (groupMod.nodes ??= {});
-		const nodeMod = (nodesMod[nodeIndex ?? this.selectedNodeInnerIndex] ??= {});
-		const typeMod = (nodeMod[section] ??= {});
-		if (typeof value === "object") {
-			const objMod = (typeMod[prop] ??= {});
-			Object.assign(objMod, value);
-		} else {
-			typeMod[prop] = value;
-		}
-	}
-
-	getEditElement(section, prop, value, placeholder, checked, checkable = true) {
-		if (value === placeholder) value = "";
-
-		const mods = this.modifications[this.selectedGroup]?.nodes?.[this.selectedNodeInnerIndex]?.[section]?.[prop];
-		if (mods) {
-			if (mods.name != null) {
-				value = mods.name;
-			}
-			if (mods.visible != null) {
-				checked = mods.visible;
-			}
-		}
-
-		return $el("div", [
-			$el("input", {
-				value,
-				placeholder,
-				type: "text",
-				onchange: (e) => {
-					this.storeModification({ section, prop, value: { name: e.target.value } });
-				},
-			}),
-			$el("label", { textContent: "Visible" }, [
-				$el("input", {
-					type: "checkbox",
-					checked,
-					disabled: !checkable,
-					onchange: (e) => {
-						this.storeModification({ section, prop, value: { visible: !!e.target.checked } });
-					},
-				}),
-			]),
-		]);
-	}
-
-	buildWidgetsPage() {
-		const widgets = this.groupData.oldToNewWidgetMap[this.selectedNodeInnerIndex];
-		const items = Object.keys(widgets ?? {});
-		const type = app.graph.extra.groupNodes[this.selectedGroup];
-		const config = type.config?.[this.selectedNodeInnerIndex]?.input;
-		this.widgetsPage.replaceChildren(
-			...items.map((oldName) => {
-				return this.getEditElement("input", oldName, widgets[oldName], oldName, config?.[oldName]?.visible !== false);
-			})
-		);
-		return !!items.length;
-	}
-
-	buildInputsPage() {
-		const inputs = this.groupData.nodeInputs[this.selectedNodeInnerIndex];
-		const items = Object.keys(inputs ?? {});
-		const type = app.graph.extra.groupNodes[this.selectedGroup];
-		const config = type.config?.[this.selectedNodeInnerIndex]?.input;
-		this.inputsPage.replaceChildren(
-			...items
-				.map((oldName) => {
-					let value = inputs[oldName];
-					if (!value) {
-						return;
-					}
-
-					return this.getEditElement("input", oldName, value, oldName, config?.[oldName]?.visible !== false);
-				})
-				.filter(Boolean)
-		);
-		return !!items.length;
-	}
-
-	buildOutputsPage() {
-		const nodes = this.groupData.nodeData.nodes;
-		const innerNodeDef = this.groupData.getNodeDef(nodes[this.selectedNodeInnerIndex]);
-		const outputs = innerNodeDef?.output ?? [];
-		const groupOutputs = this.groupData.oldToNewOutputMap[this.selectedNodeInnerIndex];
-
-		const type = app.graph.extra.groupNodes[this.selectedGroup];
-		const config = type.config?.[this.selectedNodeInnerIndex]?.output;
-		const node = this.groupData.nodeData.nodes[this.selectedNodeInnerIndex];
-		const checkable = node.type !== "PrimitiveNode";
-		this.outputsPage.replaceChildren(
-			...outputs
-				.map((type, slot) => {
-					const groupOutputIndex = groupOutputs?.[slot];
-					const oldName = innerNodeDef.output_name?.[slot] ?? type;
-					let value = config?.[slot]?.name;
-					const visible = config?.[slot]?.visible || groupOutputIndex != null;
-					if (!value || value === oldName) {
-						value = "";
-					}
-					return this.getEditElement("output", slot, value, oldName, visible, checkable);
-				})
-				.filter(Boolean)
-		);
-		return !!outputs.length;
-	}
-
-	show(type) {
-		const groupNodes = Object.keys(app.graph.extra?.groupNodes ?? {}).sort((a, b) => a.localeCompare(b));
-
-		this.innerNodesList = $el("ul.comfy-group-manage-list-items");
-		this.widgetsPage = $el("section.comfy-group-manage-node-page");
-		this.inputsPage = $el("section.comfy-group-manage-node-page");
-		this.outputsPage = $el("section.comfy-group-manage-node-page");
-		const pages = $el("div", [this.widgetsPage, this.inputsPage, this.outputsPage]);
-
-		this.tabs = [
-			["Inputs", this.inputsPage],
-			["Widgets", this.widgetsPage],
-			["Outputs", this.outputsPage],
-		].reduce((p, [name, page]) => {
-			p[name] = {
-				tab: $el("a", {
-					onclick: () => {
-						this.changeTab(name);
-					},
-					textContent: name,
-				}),
-				page,
-			};
-			return p;
-		}, {});
-
-		const outer = $el("div.comfy-group-manage-outer", [
-			$el("header", [
-				$el("h2", "Group Nodes"),
-				$el(
-					"select",
-					{
-						onchange: (e) => {
-							this.changeGroup(e.target.value);
-						},
-					},
-					groupNodes.map((g) =>
-						$el("option", {
-							textContent: g,
-							selected: "workflow/" + g === type,
-							value: g,
-						})
-					)
-				),
-			]),
-			$el("main", [
-				$el("section.comfy-group-manage-list", this.innerNodesList),
-				$el("section.comfy-group-manage-node", [
-					$el(
-						"header",
-						Object.values(this.tabs).map((t) => t.tab)
-					),
-					pages,
-				]),
-			]),
-			$el("footer", [
-				$el(
-					"button.comfy-btn",
-					{
-						onclick: (e) => {
-							const node = app.graph._nodes.find((n) => n.type === "workflow/" + this.selectedGroup);
-							if (node) {
-								alert("This group node is in use in the current workflow, please first remove these.");
-								return;
-							}
-							if (confirm(`Are you sure you want to remove the node: "${this.selectedGroup}"`)) {
-								delete app.graph.extra.groupNodes[this.selectedGroup];
-								LiteGraph.unregisterNodeType("workflow/" + this.selectedGroup);
-							}
-							this.show();
-						},
-					},
-					"Delete Group Node"
-				),
-				$el(
-					"button.comfy-btn",
-					{
-						onclick: async () => {
-							let nodesByType;
-							let recreateNodes = [];
-							const types = {};
-							for (const g in this.modifications) {
-								const type = app.graph.extra.groupNodes[g];
-								let config = (type.config ??= {});
-
-								let nodeMods = this.modifications[g]?.nodes;
-								if (nodeMods) {
-									const keys = Object.keys(nodeMods);
-									if (nodeMods[keys[0]][ORDER]) {
-										// If any node is reordered, they will all need sequencing
-										const orderedNodes = [];
-										const orderedMods = {};
-										const orderedConfig = {};
-
-										for (const n of keys) {
-											const order = nodeMods[n][ORDER].order;
-											orderedNodes[order] = type.nodes[+n];
-											orderedMods[order] = nodeMods[n];
-											orderedNodes[order].index = order;
-										}
-
-										// Rewrite links
-										for (const l of type.links) {
-											if (l[0] != null) l[0] = type.nodes[l[0]].index;
-											if (l[2] != null) l[2] = type.nodes[l[2]].index;
-										}
-
-										// Rewrite externals
-										if (type.external) {
-											for (const ext of type.external) {
-												ext[0] = type.nodes[ext[0]];
-											}
-										}
-
-										// Rewrite modifications
-										for (const id of keys) {
-											if (config[id]) {
-												orderedConfig[type.nodes[id].index] = config[id];
-											}
-											delete config[id];
-										}
-
-										type.nodes = orderedNodes;
-										nodeMods = orderedMods;
-										type.config = config = orderedConfig;
-									}
-
-									merge(config, nodeMods);
-								}
-
-								types[g] = type;
-
-								if (!nodesByType) {
-									nodesByType = app.graph._nodes.reduce((p, n) => {
-										p[n.type] ??= [];
-										p[n.type].push(n);
-										return p;
-									}, {});
-								}
-
-								const nodes = nodesByType["workflow/" + g];
-								if (nodes) recreateNodes.push(...nodes);
-							}
-
-							await GroupNodeConfig.registerFromWorkflow(types, {});
-
-							for (const node of recreateNodes) {
-								node.recreate();
-							}
-
-							this.modifications = {};
-							this.app.graph.setDirtyCanvas(true, true);
-							this.changeGroup(this.selectedGroup, false);
-						},
-					},
-					"Save"
-				),
-				$el("button.comfy-btn", { onclick: () => this.element.close() }, "Close"),
-			]),
-		]);
-
-		this.element.replaceChildren(outer);
-		this.changeGroup(type ? groupNodes.find((g) => "workflow/" + g === type) : groupNodes[0]);
-		this.element.showModal();
-
-		this.element.addEventListener("close", () => {
-			this.draggable?.dispose();
-		});
-	}
-}