Doing some experimentation

Use the WanFunInpaintToVideo node without the clip_vision_output.

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -22,7 +22,7 @@ body:
       description: Please confirm you have tried to reproduce the issue with all custom nodes disabled.
       options:
         - label: I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)
-          required: true
+          required: false
   - type: textarea
     attributes:
       label: Expected Behavior

.github/ISSUE_TEMPLATE/user-support.yml

@@ -18,7 +18,7 @@ body:
         description: Please confirm you have tried to reproduce the issue with all custom nodes disabled.
         options:
           - label: I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)
-            required: true
+            required: false
     - type: textarea
       attributes:
             label: Your question

.github/workflows/stable-release.yml

@@ -12,17 +12,17 @@ on:
         description: 'CUDA version'
         required: true
         type: string
-        default: "128"
+        default: "129"
       python_minor:
         description: 'Python minor version'
         required: true
         type: string
-        default: "12"
+        default: "13"
       python_patch:
         description: 'Python patch version'
         required: true
         type: string
-        default: "10"
+        default: "6"
 
 
 jobs:
@@ -66,8 +66,13 @@ jobs:
           curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
           ./python.exe get-pip.py
           ./python.exe -s -m pip install ../cu${{ inputs.cu }}_python_deps/*
-            sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
-            cd ..
+          sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
+
+          rm ./Lib/site-packages/torch/lib/dnnl.lib #I don't think this is actually used and I need the space
+          rm ./Lib/site-packages/torch/lib/libprotoc.lib
+          rm ./Lib/site-packages/torch/lib/libprotobuf.lib
+
+          cd ..
 
           git clone --depth 1 https://github.com/comfyanonymous/taesd
           cp taesd/*.safetensors ./ComfyUI_copy/models/vae_approx/
@@ -85,7 +90,7 @@ jobs:
 
           cd ..
 
-          "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=512m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
+          "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=768m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
           mv ComfyUI_windows_portable.7z ComfyUI/ComfyUI_windows_portable_nvidia.7z
 
           cd ComfyUI_windows_portable

.github/workflows/windows_release_dependencies.yml

@@ -17,19 +17,19 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "128"
+        default: "129"
 
       python_minor:
         description: 'python minor version'
         required: true
         type: string
-        default: "12"
+        default: "13"
 
       python_patch:
         description: 'python patch version'
         required: true
         type: string
-        default: "10"
+        default: "6"
 #  push:
 #    branches:
 #      - master

.github/workflows/windows_release_package.yml

@@ -7,19 +7,19 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "128"
+        default: "129"
 
       python_minor:
         description: 'python minor version'
         required: true
         type: string
-        default: "12"
+        default: "13"
 
       python_patch:
         description: 'python patch version'
         required: true
         type: string
-        default: "10"
+        default: "6"
 #  push:
 #    branches:
 #      - master
@@ -64,6 +64,10 @@ jobs:
             ./python.exe get-pip.py
             ./python.exe -s -m pip install ../cu${{ inputs.cu }}_python_deps/*
             sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
+
+            rm ./Lib/site-packages/torch/lib/dnnl.lib #I don't think this is actually used and I need the space
+            rm ./Lib/site-packages/torch/lib/libprotoc.lib
+            rm ./Lib/site-packages/torch/lib/libprotobuf.lib
             cd ..
 
             git clone --depth 1 https://github.com/comfyanonymous/taesd
@@ -82,7 +86,7 @@ jobs:
 
             cd ..
 
-            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=512m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
+            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=768m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
             mv ComfyUI_windows_portable.7z ComfyUI/new_ComfyUI_windows_portable_nvidia_cu${{ inputs.cu }}_or_cpu.7z
 
             cd ComfyUI_windows_portable

CODEOWNERS

@@ -5,20 +5,21 @@
 # Inlined the team members for now.
 
 # Maintainers
-*.md @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/tests/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/tests-unit/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/notebooks/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/script_examples/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/.github/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/requirements.txt @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/pyproject.toml @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+*.md @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/tests/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/tests-unit/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/notebooks/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/script_examples/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/.github/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/requirements.txt @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/pyproject.toml @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
 
 # Python web server
-/api_server/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne
-/app/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne
-/utils/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne
+/api_server/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
+/app/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
+/utils/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
 
 # Node developers
-/comfy_extras/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne
-/comfy/comfy_types/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne
+/comfy_extras/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill
+/comfy/comfy_types/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill
+/comfy_api_nodes/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill

README.md

@@ -39,7 +39,7 @@ ComfyUI lets you design and execute advanced stable diffusion pipelines using a
 ## Get Started
 
 #### [Desktop Application](https://www.comfy.org/download)
-- The easiest way to get started. 
+- The easiest way to get started.
 - Available on Windows & macOS.
 
 #### [Windows Portable Package](#installing)
@@ -65,18 +65,17 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Flux](https://comfyanonymous.github.io/ComfyUI_examples/flux/)
    - [Lumina Image 2.0](https://comfyanonymous.github.io/ComfyUI_examples/lumina2/)
    - [HiDream](https://comfyanonymous.github.io/ComfyUI_examples/hidream/)
-   - [Cosmos Predict2](https://comfyanonymous.github.io/ComfyUI_examples/cosmos_predict2/)
    - [Qwen Image](https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/)
 - Image Editing Models
    - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/)
    - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model)
    - [HiDream E1.1](https://comfyanonymous.github.io/ComfyUI_examples/hidream/#hidream-e11)
+   - [Qwen Image Edit](https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/#edit-model)
 - Video Models
    - [Stable Video Diffusion](https://comfyanonymous.github.io/ComfyUI_examples/video/)
    - [Mochi](https://comfyanonymous.github.io/ComfyUI_examples/mochi/)
    - [LTX-Video](https://comfyanonymous.github.io/ComfyUI_examples/ltxv/)
    - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/)
-   - [Nvidia Cosmos](https://comfyanonymous.github.io/ComfyUI_examples/cosmos/) and [Cosmos Predict2](https://comfyanonymous.github.io/ComfyUI_examples/cosmos_predict2/)
    - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/)
    - [Wan 2.2](https://comfyanonymous.github.io/ComfyUI_examples/wan22/)
 - Audio Models
@@ -191,7 +190,7 @@ comfy install
 
 ## Manual Install (Windows, Linux)
 
-python 3.13 is supported but using 3.12 is recommended because some custom nodes and their dependencies might not support it yet.
+Python 3.13 is very well supported. If you have trouble with some custom node dependencies you can try 3.12
 
 Git clone this repo.
 
@@ -211,27 +210,19 @@ This is the command to install the nightly with ROCm 6.4 which might have some p
 
 ### Intel GPUs (Windows and Linux)
 
-(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip (currently available in PyTorch nightly builds). More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
-  
-1. To install PyTorch nightly, use the following command:
+(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
+
+1. To install PyTorch xpu, use the following command:
+
+```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu```
+
+This is the command to install the Pytorch xpu nightly which might have some performance improvements:
 
 ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu```
 
-2. Launch ComfyUI by running `python main.py`
-
-
 (Option 2) Alternatively, Intel GPUs supported by Intel Extension for PyTorch (IPEX) can leverage IPEX for improved performance.
 
-1. For Intel® Arc™ A-Series Graphics utilizing IPEX, create a conda environment and use the commands below:
-
-```
-conda install libuv
-pip install torch==2.3.1.post0+cxx11.abi torchvision==0.18.1.post0+cxx11.abi torchaudio==2.3.1.post0+cxx11.abi intel-extension-for-pytorch==2.3.110.post0+xpu --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/ --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/
-```
-
-For other supported Intel GPUs with IPEX, visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information.
-
-Additional discussion and help can be found [here](https://github.com/comfyanonymous/ComfyUI/discussions/476).
+1. visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information.
 
 ### NVIDIA
 
@@ -352,7 +343,7 @@ Generate a self-signed certificate (not appropriate for shared/production use) a
 
 Use `--tls-keyfile key.pem --tls-certfile cert.pem` to enable TLS/SSL, the app will now be accessible with `https://...` instead of `http://...`.
 
-> Note: Windows users can use [alexisrolland/docker-openssl](https://github.com/alexisrolland/docker-openssl) or one of the [3rd party binary distributions](https://wiki.openssl.org/index.php/Binaries) to run the command example above. 
+> Note: Windows users can use [alexisrolland/docker-openssl](https://github.com/alexisrolland/docker-openssl) or one of the [3rd party binary distributions](https://wiki.openssl.org/index.php/Binaries) to run the command example above.
 <br/><br/>If you use a container, note that the volume mount `-v` can be a relative path so `... -v ".\:/openssl-certs" ...` would create the key & cert files in the current directory of your command prompt or powershell terminal.
 
 ## Support and dev channel

alembic_db/versions/e9c714da8d57_init.py

@@ -1,40 +0,0 @@
-"""init
-
-Revision ID: e9c714da8d57
-Revises:
-Create Date: 2025-05-30 20:14:33.772039
-
-"""
-from typing import Sequence, Union
-
-from alembic import op
-import sqlalchemy as sa
-
-
-# revision identifiers, used by Alembic.
-revision: str = 'e9c714da8d57'
-down_revision: Union[str, None] = None
-branch_labels: Union[str, Sequence[str], None] = None
-depends_on: Union[str, Sequence[str], None] = None
-
-
-def upgrade() -> None:
-    """Upgrade schema."""
-    op.create_table('model',
-    sa.Column('type', sa.Text(), nullable=False),
-    sa.Column('path', sa.Text(), nullable=False),
-    sa.Column('file_name', sa.Text(), nullable=True),
-    sa.Column('file_size', sa.Integer(), nullable=True),
-    sa.Column('hash', sa.Text(), nullable=True),
-    sa.Column('hash_algorithm', sa.Text(), nullable=True),
-    sa.Column('source_url', sa.Text(), nullable=True),
-    sa.Column('date_added', sa.DateTime(), server_default=sa.text('(CURRENT_TIMESTAMP)'), nullable=True),
-    sa.PrimaryKeyConstraint('type', 'path')
-    )
-
-
-def downgrade() -> None:
-    """Downgrade schema."""
-    # ### commands auto generated by Alembic - please adjust! ###
-    op.drop_table('model')
-    # ### end Alembic commands ###

app/database/models.py

@@ -1,11 +1,4 @@
-from sqlalchemy import (
-    Column,
-    Integer,
-    Text,
-    DateTime,
-)
 from sqlalchemy.orm import declarative_base
-from sqlalchemy.sql import func
 
 Base = declarative_base()
 
@@ -18,42 +11,4 @@ def to_dict(obj):
         if (val := getattr(obj, field))
     }
 
-
-class Model(Base):
-    """
-    sqlalchemy model representing a model file in the system.
-
-    This class defines the database schema for storing information about model files,
-    including their type, path, hash, and when they were added to the system.
-
-    Attributes:
-        type (Text): The type of the model, this is the name of the folder in the models folder (primary key)
-        path (Text): The file path of the model relative to the type folder (primary key)
-        file_name (Text): The name of the model file
-        file_size (Integer): The size of the model file in bytes
-        hash (Text): A hash of the model file
-        hash_algorithm (Text): The algorithm used to generate the hash
-        source_url (Text): The URL of the model file
-        date_added (DateTime): Timestamp of when the model was added to the system
-    """
-
-    __tablename__ = "model"
-
-    type = Column(Text, primary_key=True)
-    path = Column(Text, primary_key=True)
-    file_name = Column(Text)
-    file_size = Column(Integer)
-    hash = Column(Text)
-    hash_algorithm = Column(Text)
-    source_url = Column(Text)
-    date_added = Column(DateTime, server_default=func.now())
-
-    def to_dict(self):
-        """
-        Convert the model instance to a dictionary representation.
-
-        Returns:
-            dict: A dictionary containing the attributes of the model
-        """
-        dict = to_dict(self)
-        return dict
+# TODO: Define models here

app/frontend_management.py

@@ -196,17 +196,6 @@ def download_release_asset_zip(release: Release, destination_path: str) -> None:
 
 
 class FrontendManager:
-    """
-    A class to manage ComfyUI frontend versions and installations.
-
-    This class handles the initialization and management of different frontend versions,
-    including the default frontend from the pip package and custom frontend versions
-    from GitHub repositories.
-
-    Attributes:
-        CUSTOM_FRONTENDS_ROOT (str): The root directory where custom frontend versions are stored.
-    """
-
     CUSTOM_FRONTENDS_ROOT = str(Path(__file__).parents[1] / "web_custom_versions")
 
     @classmethod
@@ -216,15 +205,6 @@ class FrontendManager:
 
     @classmethod
     def default_frontend_path(cls) -> str:
-        """
-        Get the path to the default frontend installation from the pip package.
-
-        Returns:
-            str: The path to the default frontend static files.
-
-        Raises:
-            SystemExit: If the comfyui-frontend-package is not installed.
-        """
         try:
             import comfyui_frontend_package
 
@@ -245,15 +225,6 @@ comfyui-frontend-package is not installed.
 
     @classmethod
     def templates_path(cls) -> str:
-        """
-        Get the path to the workflow templates.
-
-        Returns:
-            str: The path to the workflow templates directory.
-
-        Raises:
-            SystemExit: If the comfyui-workflow-templates package is not installed.
-        """
         try:
             import comfyui_workflow_templates
 
@@ -289,16 +260,11 @@ comfyui-workflow-templates is not installed.
     @classmethod
     def parse_version_string(cls, value: str) -> tuple[str, str, str]:
         """
-        Parse a version string into its components.
-
-        The version string should be in the format: 'owner/repo@version'
-        where version can be either a semantic version (v1.2.3) or 'latest'.
-
         Args:
             value (str): The version string to parse.
 
         Returns:
-            tuple[str, str, str]: A tuple containing (owner, repo, version).
+            tuple[str, str]: A tuple containing provider name and version.
 
         Raises:
             argparse.ArgumentTypeError: If the version string is invalid.
@@ -315,22 +281,18 @@ comfyui-workflow-templates is not installed.
         cls, version_string: str, provider: Optional[FrontEndProvider] = None
     ) -> str:
         """
-        Initialize a frontend version without error handling.
-
-        This method attempts to initialize a specific frontend version, either from
-        the default pip package or from a custom GitHub repository. It will download
-        and extract the frontend files if necessary.
+        Initializes the frontend for the specified version.
 
         Args:
-            version_string (str): The version string specifying which frontend to use.
-            provider (FrontEndProvider, optional): The provider to use for custom frontends.
+            version_string (str): The version string.
+            provider (FrontEndProvider, optional): The provider to use. Defaults to None.
 
         Returns:
             str: The path to the initialized frontend.
 
         Raises:
-            Exception: If there is an error during initialization (e.g., network timeout,
-                      invalid URL, or missing assets).
+            Exception: If there is an error during the initialization process.
+            main error source might be request timeout or invalid URL.
         """
         if version_string == DEFAULT_VERSION_STRING:
             check_frontend_version()
@@ -382,17 +344,13 @@ comfyui-workflow-templates is not installed.
     @classmethod
     def init_frontend(cls, version_string: str) -> str:
         """
-        Initialize a frontend version with error handling.
-
-        This is the main method to initialize a frontend version. It wraps init_frontend_unsafe
-        with error handling, falling back to the default frontend if initialization fails.
+        Initializes the frontend with the specified version string.
 
         Args:
-            version_string (str): The version string specifying which frontend to use.
+            version_string (str): The version string to initialize the frontend with.
 
         Returns:
-            str: The path to the initialized frontend. If initialization fails,
-                 returns the path to the default frontend.
+            str: The path of the initialized frontend.
         """
         try:
             return cls.init_frontend_unsafe(version_string)

app/model_processor.py

@@ -1,331 +0,0 @@
-import os
-import logging
-import time
-
-import requests
-from tqdm import tqdm
-from folder_paths import get_relative_path, get_full_path
-from app.database.db import create_session, dependencies_available, can_create_session
-import blake3
-import comfy.utils
-
-
-if dependencies_available():
-    from app.database.models import Model
-
-
-class ModelProcessor:
-    def _validate_path(self, model_path):
-        try:
-            if not self._file_exists(model_path):
-                logging.error(f"Model file not found: {model_path}")
-                return None
-
-            result = get_relative_path(model_path)
-            if not result:
-                logging.error(
-                    f"Model file not in a recognized model directory: {model_path}"
-                )
-                return None
-
-            return result
-        except Exception as e:
-            logging.error(f"Error validating model path {model_path}: {str(e)}")
-            return None
-
-    def _file_exists(self, path):
-        """Check if a file exists."""
-        return os.path.exists(path)
-
-    def _get_file_size(self, path):
-        """Get file size."""
-        return os.path.getsize(path)
-
-    def _get_hasher(self):
-        return blake3.blake3()
-
-    def _hash_file(self, model_path):
-        try:
-            hasher = self._get_hasher()
-            with open(model_path, "rb", buffering=0) as f:
-                b = bytearray(128 * 1024)
-                mv = memoryview(b)
-                while n := f.readinto(mv):
-                    hasher.update(mv[:n])
-            return hasher.hexdigest()
-        except Exception as e:
-            logging.error(f"Error hashing file {model_path}: {str(e)}")
-            return None
-
-    def _get_existing_model(self, session, model_type, model_relative_path):
-        return (
-            session.query(Model)
-            .filter(Model.type == model_type)
-            .filter(Model.path == model_relative_path)
-            .first()
-        )
-
-    def _ensure_source_url(self, session, model, source_url):
-        if model.source_url is None:
-            model.source_url = source_url
-            session.commit()
-
-    def _update_database(
-        self,
-        session,
-        model_type,
-        model_path,
-        model_relative_path,
-        model_hash,
-        model,
-        source_url,
-    ):
-        try:
-            if not model:
-                model = self._get_existing_model(
-                    session, model_type, model_relative_path
-                )
-
-            if not model:
-                model = Model(
-                    path=model_relative_path,
-                    type=model_type,
-                    file_name=os.path.basename(model_path),
-                )
-                session.add(model)
-
-            model.file_size = self._get_file_size(model_path)
-            model.hash = model_hash
-            if model_hash:
-                model.hash_algorithm = "blake3"
-            model.source_url = source_url
-
-            session.commit()
-            return model
-        except Exception as e:
-            logging.error(
-                f"Error updating database for {model_relative_path}: {str(e)}"
-            )
-
-    def process_file(self, model_path, source_url=None, model_hash=None):
-        """
-        Process a model file and update the database with metadata.
-        If the file already exists and matches the database, it will not be processed again.
-        Returns the model object or if an error occurs, returns None.
-        """
-        try:
-            if not can_create_session():
-                return
-
-            result = self._validate_path(model_path)
-            if not result:
-                return
-            model_type, model_relative_path = result
-
-            with create_session() as session:
-                session.expire_on_commit = False
-
-                existing_model = self._get_existing_model(
-                    session, model_type, model_relative_path
-                )
-                if (
-                    existing_model
-                    and existing_model.hash
-                    and existing_model.file_size == self._get_file_size(model_path)
-                ):
-                    # File exists with hash and same size, no need to process
-                    self._ensure_source_url(session, existing_model, source_url)
-                    return existing_model
-
-                if model_hash:
-                    model_hash = model_hash.lower()
-                    logging.info(f"Using provided hash: {model_hash}")
-                else:
-                    start_time = time.time()
-                    logging.info(f"Hashing model {model_relative_path}")
-                    model_hash = self._hash_file(model_path)
-                    if not model_hash:
-                        return
-                    logging.info(
-                        f"Model hash: {model_hash} (duration: {time.time() - start_time} seconds)"
-                    )
-
-                return self._update_database(
-                    session,
-                    model_type,
-                    model_path,
-                    model_relative_path,
-                    model_hash,
-                    existing_model,
-                    source_url,
-                )
-        except Exception as e:
-            logging.error(f"Error processing model file {model_path}: {str(e)}")
-            return None
-
-    def retrieve_model_by_hash(self, model_hash, model_type=None, session=None):
-        """
-        Retrieve a model file from the database by hash and optionally by model type.
-        Returns the model object or None if the model doesnt exist or an error occurs.
-        """
-        try:
-            if not can_create_session():
-                return
-
-            dispose_session = False
-
-            if session is None:
-                session = create_session()
-                dispose_session = True
-
-            model = session.query(Model).filter(Model.hash == model_hash)
-            if model_type is not None:
-                model = model.filter(Model.type == model_type)
-            return model.first()
-        except Exception as e:
-            logging.error(f"Error retrieving model by hash {model_hash}: {str(e)}")
-            return None
-        finally:
-            if dispose_session:
-                session.close()
-
-    def retrieve_hash(self, model_path, model_type=None):
-        """
-        Retrieve the hash of a model file from the database.
-        Returns the hash or None if the model doesnt exist or an error occurs.
-        """
-        try:
-            if not can_create_session():
-                return
-
-            if model_type is not None:
-                result = self._validate_path(model_path)
-                if not result:
-                    return None
-            model_type, model_relative_path = result
-
-            with create_session() as session:
-                model = self._get_existing_model(
-                    session, model_type, model_relative_path
-                )
-                if model and model.hash:
-                    return model.hash
-                return None
-        except Exception as e:
-            logging.error(f"Error retrieving hash for {model_path}: {str(e)}")
-            return None
-
-    def _validate_file_extension(self, file_name):
-        """Validate that the file extension is supported."""
-        extension = os.path.splitext(file_name)[1]
-        if extension not in (".safetensors", ".sft", ".txt", ".csv", ".json", ".yaml"):
-            raise ValueError(f"Unsupported unsafe file for download: {file_name}")
-
-    def _check_existing_file(self, model_type, file_name, expected_hash):
-        """Check if file exists and has correct hash."""
-        destination_path = get_full_path(model_type, file_name, allow_missing=True)
-        if self._file_exists(destination_path):
-            model = self.process_file(destination_path)
-            if model and (expected_hash is None or model.hash == expected_hash):
-                logging.debug(
-                    f"File {destination_path} already exists in the database and has the correct hash or no hash was provided."
-                )
-                return destination_path
-            else:
-                raise ValueError(
-                    f"File {destination_path} exists with hash {model.hash if model else 'unknown'} but expected {expected_hash}. Please delete the file and try again."
-                )
-        return None
-
-    def _check_existing_file_by_hash(self, hash, type, url):
-        """Check if a file with the given hash exists in the database and on disk."""
-        hash = hash.lower()
-        with create_session() as session:
-            model = self.retrieve_model_by_hash(hash, type, session)
-            if model:
-                existing_path = get_full_path(type, model.path)
-                if existing_path:
-                    logging.debug(
-                        f"File {model.path} already exists in the database at {existing_path}"
-                    )
-                    self._ensure_source_url(session, model, url)
-                    return existing_path
-                else:
-                    logging.debug(
-                        f"File {model.path} exists in the database but not on disk"
-                    )
-        return None
-
-    def _download_file(self, url, destination_path, hasher):
-        """Download a file and update the hasher with its contents."""
-        response = requests.get(url, stream=True)
-        logging.info(f"Downloading {url} to {destination_path}")
-
-        with open(destination_path, "wb") as f:
-            total_size = int(response.headers.get("content-length", 0))
-            if total_size > 0:
-                pbar = comfy.utils.ProgressBar(total_size)
-            else:
-                pbar = None
-            with tqdm(total=total_size, unit="B", unit_scale=True) as progress_bar:
-                for chunk in response.iter_content(chunk_size=128 * 1024):
-                    if chunk:
-                        f.write(chunk)
-                        hasher.update(chunk)
-                        progress_bar.update(len(chunk))
-                        if pbar:
-                            pbar.update(len(chunk))
-
-    def _verify_downloaded_hash(self, calculated_hash, expected_hash, destination_path):
-        """Verify that the downloaded file has the expected hash."""
-        if expected_hash is not None and calculated_hash != expected_hash:
-            self._remove_file(destination_path)
-            raise ValueError(
-                f"Downloaded file hash {calculated_hash} does not match expected hash {expected_hash}"
-            )
-
-    def _remove_file(self, file_path):
-        """Remove a file from disk."""
-        os.remove(file_path)
-
-    def ensure_downloaded(self, type, url, desired_file_name, hash=None):
-        """
-        Ensure a model file is downloaded and has the correct hash.
-        Returns the path to the downloaded file.
-        """
-        logging.debug(
-            f"Ensuring {type} file is downloaded. URL='{url}' Destination='{desired_file_name}' Hash='{hash}'"
-        )
-
-        # Validate file extension
-        self._validate_file_extension(desired_file_name)
-
-        # Check if file exists with correct hash
-        if hash:
-            existing_path = self._check_existing_file_by_hash(hash, type, url)
-            if existing_path:
-                return existing_path
-
-        # Check if file exists locally
-        destination_path = get_full_path(type, desired_file_name, allow_missing=True)
-        existing_path = self._check_existing_file(type, desired_file_name, hash)
-        if existing_path:
-            return existing_path
-
-        # Download the file
-        hasher = self._get_hasher()
-        self._download_file(url, destination_path, hasher)
-
-        # Verify hash
-        calculated_hash = hasher.hexdigest()
-        self._verify_downloaded_hash(calculated_hash, hash, destination_path)
-
-        # Update database
-        self.process_file(destination_path, url, calculated_hash)
-
-        # TODO: Notify frontend to reload models
-
-        return destination_path
-
-
-model_processor = ModelProcessor()

app/user_manager.py

@@ -363,10 +363,17 @@ class UserManager():
             if not overwrite and os.path.exists(path):
                 return web.Response(status=409, text="File already exists")
 
-            body = await request.read()
+            try:
+                body = await request.read()
 
-            with open(path, "wb") as f:
-                f.write(body)
+                with open(path, "wb") as f:
+                    f.write(body)
+            except OSError as e:
+                logging.warning(f"Error saving file '{path}': {e}")
+                return web.Response(
+                    status=400,
+                    reason="Invalid filename. Please avoid special characters like :\\/*?\"<>|"
+                )
 
             user_path = self.get_request_user_filepath(request, None)
             if full_info:

comfy/audio_encoders/audio_encoders.py

@@ -0,0 +1,42 @@
+from .wav2vec2 import Wav2Vec2Model
+import comfy.model_management
+import comfy.ops
+import comfy.utils
+import logging
+import torchaudio
+
+
+class AudioEncoderModel():
+    def __init__(self, config):
+        self.load_device = comfy.model_management.text_encoder_device()
+        offload_device = comfy.model_management.text_encoder_offload_device()
+        self.dtype = comfy.model_management.text_encoder_dtype(self.load_device)
+        self.model = Wav2Vec2Model(dtype=self.dtype, device=offload_device, operations=comfy.ops.manual_cast)
+        self.model.eval()
+        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.model_sample_rate = 16000
+
+    def load_sd(self, sd):
+        return self.model.load_state_dict(sd, strict=False)
+
+    def get_sd(self):
+        return self.model.state_dict()
+
+    def encode_audio(self, audio, sample_rate):
+        comfy.model_management.load_model_gpu(self.patcher)
+        audio = torchaudio.functional.resample(audio, sample_rate, self.model_sample_rate)
+        out, all_layers = self.model(audio.to(self.load_device))
+        outputs = {}
+        outputs["encoded_audio"] = out
+        outputs["encoded_audio_all_layers"] = all_layers
+        return outputs
+
+
+def load_audio_encoder_from_sd(sd, prefix=""):
+    audio_encoder = AudioEncoderModel(None)
+    sd = comfy.utils.state_dict_prefix_replace(sd, {"wav2vec2.": ""})
+    m, u = audio_encoder.load_sd(sd)
+    if len(m) > 0:
+        logging.warning("missing audio encoder: {}".format(m))
+
+    return audio_encoder

comfy/audio_encoders/wav2vec2.py

@@ -0,0 +1,207 @@
+import torch
+import torch.nn as nn
+from comfy.ldm.modules.attention import optimized_attention_masked
+
+
+class LayerNormConv(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, stride, bias=False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.conv = operations.Conv1d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, bias=bias, device=device, dtype=dtype)
+        self.layer_norm = operations.LayerNorm(out_channels, elementwise_affine=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return torch.nn.functional.gelu(self.layer_norm(x.transpose(-2, -1)).transpose(-2, -1))
+
+
+class ConvFeatureEncoder(nn.Module):
+    def __init__(self, conv_dim, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.conv_layers = nn.ModuleList([
+            LayerNormConv(1, conv_dim, kernel_size=10, stride=5, bias=True, device=device, dtype=dtype, operations=operations),
+            LayerNormConv(conv_dim, conv_dim, kernel_size=3, stride=2, bias=True, device=device, dtype=dtype, operations=operations),
+            LayerNormConv(conv_dim, conv_dim, kernel_size=3, stride=2, bias=True, device=device, dtype=dtype, operations=operations),
+            LayerNormConv(conv_dim, conv_dim, kernel_size=3, stride=2, bias=True, device=device, dtype=dtype, operations=operations),
+            LayerNormConv(conv_dim, conv_dim, kernel_size=3, stride=2, bias=True, device=device, dtype=dtype, operations=operations),
+            LayerNormConv(conv_dim, conv_dim, kernel_size=2, stride=2, bias=True, device=device, dtype=dtype, operations=operations),
+            LayerNormConv(conv_dim, conv_dim, kernel_size=2, stride=2, bias=True, device=device, dtype=dtype, operations=operations),
+        ])
+
+    def forward(self, x):
+        x = x.unsqueeze(1)
+
+        for conv in self.conv_layers:
+            x = conv(x)
+
+        return x.transpose(1, 2)
+
+
+class FeatureProjection(nn.Module):
+    def __init__(self, conv_dim, embed_dim, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.layer_norm = operations.LayerNorm(conv_dim, eps=1e-05, device=device, dtype=dtype)
+        self.projection = operations.Linear(conv_dim, embed_dim, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.layer_norm(x)
+        x = self.projection(x)
+        return x
+
+
+class PositionalConvEmbedding(nn.Module):
+    def __init__(self, embed_dim=768, kernel_size=128, groups=16):
+        super().__init__()
+        self.conv = nn.Conv1d(
+            embed_dim,
+            embed_dim,
+            kernel_size=kernel_size,
+            padding=kernel_size // 2,
+            groups=groups,
+        )
+        self.conv = torch.nn.utils.parametrizations.weight_norm(self.conv, name="weight", dim=2)
+        self.activation = nn.GELU()
+
+    def forward(self, x):
+        x = x.transpose(1, 2)
+        x = self.conv(x)[:, :, :-1]
+        x = self.activation(x)
+        x = x.transpose(1, 2)
+        return x
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(
+        self,
+        embed_dim=768,
+        num_heads=12,
+        num_layers=12,
+        mlp_ratio=4.0,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+
+        self.pos_conv_embed = PositionalConvEmbedding(embed_dim=embed_dim)
+        self.layers = nn.ModuleList([
+            TransformerEncoderLayer(
+                embed_dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                device=device, dtype=dtype, operations=operations
+            )
+            for _ in range(num_layers)
+        ])
+
+        self.layer_norm = operations.LayerNorm(embed_dim, eps=1e-05, device=device, dtype=dtype)
+
+    def forward(self, x, mask=None):
+        x = x + self.pos_conv_embed(x)
+        all_x = ()
+        for layer in self.layers:
+            all_x += (x,)
+            x = layer(x, mask)
+        x = self.layer_norm(x)
+        all_x += (x,)
+        return x, all_x
+
+
+class Attention(nn.Module):
+    def __init__(self, embed_dim, num_heads, bias=True, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.head_dim = embed_dim // num_heads
+
+        self.k_proj = operations.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+        self.v_proj = operations.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+        self.q_proj = operations.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+        self.out_proj = operations.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+
+    def forward(self, x, mask=None):
+        assert (mask is None)  # TODO?
+        q = self.q_proj(x)
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+
+        out = optimized_attention_masked(q, k, v, self.num_heads)
+        return self.out_proj(out)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, embed_dim, mlp_ratio, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.intermediate_dense = operations.Linear(embed_dim, int(embed_dim * mlp_ratio), device=device, dtype=dtype)
+        self.output_dense = operations.Linear(int(embed_dim * mlp_ratio), embed_dim, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.intermediate_dense(x)
+        x = torch.nn.functional.gelu(x)
+        x = self.output_dense(x)
+        return x
+
+
+class TransformerEncoderLayer(nn.Module):
+    def __init__(
+        self,
+        embed_dim=768,
+        num_heads=12,
+        mlp_ratio=4.0,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+
+        self.attention = Attention(embed_dim, num_heads, device=device, dtype=dtype, operations=operations)
+
+        self.layer_norm = operations.LayerNorm(embed_dim, device=device, dtype=dtype)
+        self.feed_forward = FeedForward(embed_dim, mlp_ratio, device=device, dtype=dtype, operations=operations)
+        self.final_layer_norm = operations.LayerNorm(embed_dim, device=device, dtype=dtype)
+
+    def forward(self, x, mask=None):
+        residual = x
+        x = self.layer_norm(x)
+        x = self.attention(x, mask=mask)
+        x = residual + x
+
+        x = x + self.feed_forward(self.final_layer_norm(x))
+        return x
+
+
+class Wav2Vec2Model(nn.Module):
+    """Complete Wav2Vec 2.0 model."""
+
+    def __init__(
+        self,
+        embed_dim=1024,
+        final_dim=256,
+        num_heads=16,
+        num_layers=24,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+
+        conv_dim = 512
+        self.feature_extractor = ConvFeatureEncoder(conv_dim, device=device, dtype=dtype, operations=operations)
+        self.feature_projection = FeatureProjection(conv_dim, embed_dim, device=device, dtype=dtype, operations=operations)
+
+        self.masked_spec_embed = nn.Parameter(torch.empty(embed_dim, device=device, dtype=dtype))
+
+        self.encoder = TransformerEncoder(
+            embed_dim=embed_dim,
+            num_heads=num_heads,
+            num_layers=num_layers,
+            device=device, dtype=dtype, operations=operations
+        )
+
+    def forward(self, x, mask_time_indices=None, return_dict=False):
+
+        x = torch.mean(x, dim=1)
+
+        x = (x - x.mean()) / torch.sqrt(x.var() + 1e-7)
+
+        features = self.feature_extractor(x)
+        features = self.feature_projection(features)
+
+        batch_size, seq_len, _ = features.shape
+
+        x, all_x = self.encoder(features)
+
+        return x, all_x

comfy/cli_args.py

@@ -132,6 +132,8 @@ parser.add_argument("--reserve-vram", type=float, default=None, help="Set the am
 
 parser.add_argument("--async-offload", action="store_true", help="Use async weight offloading.")
 
+parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.")
+
 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.")
@@ -210,7 +212,6 @@ database_default_path = os.path.abspath(
     os.path.join(os.path.dirname(__file__), "..", "user", "comfyui.db")
 )
 parser.add_argument("--database-url", type=str, default=f"sqlite:///{database_default_path}", help="Specify the database URL, e.g. for an in-memory database you can use 'sqlite:///:memory:'.")
-parser.add_argument("--disable-model-processing", action="store_true", help="Disable model file processing, e.g. computing hashes and extracting metadata.")
 
 if comfy.options.args_parsing:
     args = parser.parse_args()

comfy/clip_model.py

@@ -97,7 +97,7 @@ class CLIPTextModel_(torch.nn.Module):
         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)
 
-    def forward(self, input_tokens=None, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32):
+    def forward(self, input_tokens=None, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32, embeds_info=[]):
         if embeds is not None:
             x = embeds + comfy.ops.cast_to(self.embeddings.position_embedding.weight, dtype=dtype, device=embeds.device)
         else:

comfy/context_windows.py

@@ -0,0 +1,540 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING, Callable
+import torch
+import numpy as np
+import collections
+from dataclasses import dataclass
+from abc import ABC, abstractmethod
+import logging
+import comfy.model_management
+import comfy.patcher_extension
+if TYPE_CHECKING:
+    from comfy.model_base import BaseModel
+    from comfy.model_patcher import ModelPatcher
+    from comfy.controlnet import ControlBase
+
+
+class ContextWindowABC(ABC):
+    def __init__(self):
+        ...
+
+    @abstractmethod
+    def get_tensor(self, full: torch.Tensor) -> torch.Tensor:
+        """
+        Get torch.Tensor applicable to current window.
+        """
+        raise NotImplementedError("Not implemented.")
+
+    @abstractmethod
+    def add_window(self, full: torch.Tensor, to_add: torch.Tensor) -> torch.Tensor:
+        """
+        Apply torch.Tensor of window to the full tensor, in place. Returns reference to updated full tensor, not a copy.
+        """
+        raise NotImplementedError("Not implemented.")
+
+class ContextHandlerABC(ABC):
+    def __init__(self):
+        ...
+
+    @abstractmethod
+    def should_use_context(self, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]) -> bool:
+        raise NotImplementedError("Not implemented.")
+
+    @abstractmethod
+    def get_resized_cond(self, cond_in: list[dict], x_in: torch.Tensor, window: ContextWindowABC, device=None) -> list:
+        raise NotImplementedError("Not implemented.")
+
+    @abstractmethod
+    def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+        raise NotImplementedError("Not implemented.")
+
+
+
+class IndexListContextWindow(ContextWindowABC):
+    def __init__(self, index_list: list[int], dim: int=0):
+        self.index_list = index_list
+        self.context_length = len(index_list)
+        self.dim = dim
+
+    def get_tensor(self, full: torch.Tensor, device=None, dim=None) -> torch.Tensor:
+        if dim is None:
+            dim = self.dim
+        if dim == 0 and full.shape[dim] == 1:
+            return full
+        idx = [slice(None)] * dim + [self.index_list]
+        return full[idx].to(device)
+
+    def add_window(self, full: torch.Tensor, to_add: torch.Tensor, dim=None) -> torch.Tensor:
+        if dim is None:
+            dim = self.dim
+        idx = [slice(None)] * dim + [self.index_list]
+        full[idx] += to_add
+        return full
+
+
+class IndexListCallbacks:
+    EVALUATE_CONTEXT_WINDOWS = "evaluate_context_windows"
+    COMBINE_CONTEXT_WINDOW_RESULTS = "combine_context_window_results"
+    EXECUTE_START = "execute_start"
+    EXECUTE_CLEANUP = "execute_cleanup"
+
+    def init_callbacks(self):
+        return {}
+
+
+@dataclass
+class ContextSchedule:
+    name: str
+    func: Callable
+
+@dataclass
+class ContextFuseMethod:
+    name: str
+    func: Callable
+
+ContextResults = collections.namedtuple("ContextResults", ['window_idx', 'sub_conds_out', 'sub_conds', 'window'])
+class IndexListContextHandler(ContextHandlerABC):
+    def __init__(self, context_schedule: ContextSchedule, fuse_method: ContextFuseMethod, context_length: int=1, context_overlap: int=0, context_stride: int=1, closed_loop=False, dim=0):
+        self.context_schedule = context_schedule
+        self.fuse_method = fuse_method
+        self.context_length = context_length
+        self.context_overlap = context_overlap
+        self.context_stride = context_stride
+        self.closed_loop = closed_loop
+        self.dim = dim
+        self._step = 0
+
+        self.callbacks = {}
+
+    def should_use_context(self, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]) -> bool:
+        # for now, assume first dim is batch - should have stored on BaseModel in actual implementation
+        if x_in.size(self.dim) > self.context_length:
+            logging.info(f"Using context windows {self.context_length} for {x_in.size(self.dim)} frames.")
+            return True
+        return False
+
+    def prepare_control_objects(self, control: ControlBase, device=None) -> ControlBase:
+        if control.previous_controlnet is not None:
+            self.prepare_control_objects(control.previous_controlnet, device)
+        return control
+
+    def get_resized_cond(self, cond_in: list[dict], x_in: torch.Tensor, window: IndexListContextWindow, device=None) -> list:
+        if cond_in is None:
+            return None
+        # reuse or resize cond items to match context requirements
+        resized_cond = []
+        # cond object is a list containing a dict - outer list is irrelevant, so just loop through it
+        for actual_cond in cond_in:
+            resized_actual_cond = actual_cond.copy()
+            # now we are in the inner dict - "pooled_output" is a tensor, "control" is a ControlBase object, "model_conds" is dictionary
+            for key in actual_cond:
+                try:
+                    cond_item = actual_cond[key]
+                    if isinstance(cond_item, torch.Tensor):
+                        # check that tensor is the expected length - x.size(0)
+                        if self.dim < cond_item.ndim and cond_item.size(self.dim) == x_in.size(self.dim):
+                            # if so, it's subsetting time - tell controls the expected indeces so they can handle them
+                            actual_cond_item = window.get_tensor(cond_item)
+                            resized_actual_cond[key] = actual_cond_item.to(device)
+                        else:
+                            resized_actual_cond[key] = cond_item.to(device)
+                    # look for control
+                    elif key == "control":
+                        resized_actual_cond[key] = self.prepare_control_objects(cond_item, device)
+                    elif isinstance(cond_item, dict):
+                        new_cond_item = cond_item.copy()
+                        # when in dictionary, look for tensors and CONDCrossAttn [comfy/conds.py] (has cond attr that is a tensor)
+                        for cond_key, cond_value in new_cond_item.items():
+                            if isinstance(cond_value, torch.Tensor):
+                                if cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim):
+                                    new_cond_item[cond_key] = window.get_tensor(cond_value, device)
+                            # if has cond that is a Tensor, check if needs to be subset
+                            elif hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
+                                if cond_value.cond.ndim < self.dim and cond_value.cond.size(0) == x_in.size(self.dim):
+                                    new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(cond_value.cond, device))
+                            elif cond_key == "num_video_frames": # for SVD
+                                new_cond_item[cond_key] = cond_value._copy_with(cond_value.cond)
+                                new_cond_item[cond_key].cond = window.context_length
+                        resized_actual_cond[key] = new_cond_item
+                    else:
+                        resized_actual_cond[key] = cond_item
+                finally:
+                    del cond_item  # just in case to prevent VRAM issues
+            resized_cond.append(resized_actual_cond)
+        return resized_cond
+
+    def set_step(self, timestep: torch.Tensor, model_options: dict[str]):
+        mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep, rtol=0.0001)
+        matches = torch.nonzero(mask)
+        if torch.numel(matches) == 0:
+            raise Exception("No sample_sigmas matched current timestep; something went wrong.")
+        self._step = int(matches[0].item())
+
+    def get_context_windows(self, model: BaseModel, x_in: torch.Tensor, model_options: dict[str]) -> list[IndexListContextWindow]:
+        full_length = x_in.size(self.dim) # TODO: choose dim based on model
+        context_windows = self.context_schedule.func(full_length, self, model_options)
+        context_windows = [IndexListContextWindow(window, dim=self.dim) for window in context_windows]
+        return context_windows
+
+    def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+        self.set_step(timestep, model_options)
+        context_windows = self.get_context_windows(model, x_in, model_options)
+        enumerated_context_windows = list(enumerate(context_windows))
+
+        conds_final = [torch.zeros_like(x_in) for _ in conds]
+        if self.fuse_method.name == ContextFuseMethods.RELATIVE:
+            counts_final = [torch.ones(get_shape_for_dim(x_in, self.dim), device=x_in.device) for _ in conds]
+        else:
+            counts_final = [torch.zeros(get_shape_for_dim(x_in, self.dim), device=x_in.device) for _ in conds]
+        biases_final = [([0.0] * x_in.shape[self.dim]) for _ in conds]
+
+        for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.EXECUTE_START, self.callbacks):
+            callback(self, model, x_in, conds, timestep, model_options)
+
+        for enum_window in enumerated_context_windows:
+            results = self.evaluate_context_windows(calc_cond_batch, model, x_in, conds, timestep, [enum_window], model_options)
+            for result in results:
+                self.combine_context_window_results(x_in, result.sub_conds_out, result.sub_conds, result.window, result.window_idx, len(enumerated_context_windows), timestep,
+                                            conds_final, counts_final, biases_final)
+        try:
+            # finalize conds
+            if self.fuse_method.name == ContextFuseMethods.RELATIVE:
+                # relative is already normalized, so return as is
+                del counts_final
+                return conds_final
+            else:
+                # normalize conds via division by context usage counts
+                for i in range(len(conds_final)):
+                    conds_final[i] /= counts_final[i]
+                del counts_final
+                return conds_final
+        finally:
+            for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.EXECUTE_CLEANUP, self.callbacks):
+                callback(self, model, x_in, conds, timestep, model_options)
+
+    def evaluate_context_windows(self, calc_cond_batch: Callable, model: BaseModel, x_in: torch.Tensor, conds, timestep: torch.Tensor, enumerated_context_windows: list[tuple[int, IndexListContextWindow]],
+                                model_options, device=None, first_device=None):
+        results: list[ContextResults] = []
+        for window_idx, window in enumerated_context_windows:
+            # allow processing to end between context window executions for faster Cancel
+            comfy.model_management.throw_exception_if_processing_interrupted()
+
+            for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.EVALUATE_CONTEXT_WINDOWS, self.callbacks):
+                callback(self, model, x_in, conds, timestep, model_options, window_idx, window, model_options, device, first_device)
+
+            # update exposed params
+            model_options["transformer_options"]["context_window"] = window
+            # get subsections of x, timestep, conds
+            sub_x = window.get_tensor(x_in, device)
+            sub_timestep = window.get_tensor(timestep, device, dim=0)
+            sub_conds = [self.get_resized_cond(cond, x_in, window, device) for cond in conds]
+
+            sub_conds_out = calc_cond_batch(model, sub_conds, sub_x, sub_timestep, model_options)
+            if device is not None:
+                for i in range(len(sub_conds_out)):
+                    sub_conds_out[i] = sub_conds_out[i].to(x_in.device)
+            results.append(ContextResults(window_idx, sub_conds_out, sub_conds, window))
+        return results
+
+
+    def combine_context_window_results(self, x_in: torch.Tensor, sub_conds_out, sub_conds, window: IndexListContextWindow, window_idx: int, total_windows: int, timestep: torch.Tensor,
+                                    conds_final: list[torch.Tensor], counts_final: list[torch.Tensor], biases_final: list[torch.Tensor]):
+        if self.fuse_method.name == ContextFuseMethods.RELATIVE:
+            for pos, idx in enumerate(window.index_list):
+                # bias is the influence of a specific index in relation to the whole context window
+                bias = 1 - abs(idx - (window.index_list[0] + window.index_list[-1]) / 2) / ((window.index_list[-1] - window.index_list[0] + 1e-2) / 2)
+                bias = max(1e-2, bias)
+                # take weighted average relative to total bias of current idx
+                for i in range(len(sub_conds_out)):
+                    bias_total = biases_final[i][idx]
+                    prev_weight = (bias_total / (bias_total + bias))
+                    new_weight = (bias / (bias_total + bias))
+                    # account for dims of tensors
+                    idx_window = [slice(None)] * self.dim + [idx]
+                    pos_window = [slice(None)] * self.dim + [pos]
+                    # apply new values
+                    conds_final[i][idx_window] = conds_final[i][idx_window] * prev_weight + sub_conds_out[i][pos_window] * new_weight
+                    biases_final[i][idx] = bias_total + bias
+        else:
+            # add conds and counts based on weights of fuse method
+            weights = get_context_weights(window.context_length, x_in.shape[self.dim], window.index_list, self, sigma=timestep)
+            weights_tensor = match_weights_to_dim(weights, x_in, self.dim, device=x_in.device)
+            for i in range(len(sub_conds_out)):
+                window.add_window(conds_final[i], sub_conds_out[i] * weights_tensor)
+                window.add_window(counts_final[i], weights_tensor)
+
+        for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.COMBINE_CONTEXT_WINDOW_RESULTS, self.callbacks):
+            callback(self, x_in, sub_conds_out, sub_conds, window, window_idx, total_windows, timestep, conds_final, counts_final, biases_final)
+
+
+def _prepare_sampling_wrapper(executor, model, noise_shape: torch.Tensor, *args, **kwargs):
+    # limit noise_shape length to context_length for more accurate vram use estimation
+    model_options = kwargs.get("model_options", None)
+    if model_options is None:
+        raise Exception("model_options not found in prepare_sampling_wrapper; this should never happen, something went wrong.")
+    handler: IndexListContextHandler = model_options.get("context_handler", None)
+    if handler is not None:
+        noise_shape = list(noise_shape)
+        noise_shape[handler.dim] = min(noise_shape[handler.dim], handler.context_length)
+    return executor(model, noise_shape, *args, **kwargs)
+
+
+def create_prepare_sampling_wrapper(model: ModelPatcher):
+    model.add_wrapper_with_key(
+        comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING,
+        "ContextWindows_prepare_sampling",
+        _prepare_sampling_wrapper
+    )
+
+
+def match_weights_to_dim(weights: list[float], x_in: torch.Tensor, dim: int, device=None) -> torch.Tensor:
+    total_dims = len(x_in.shape)
+    weights_tensor = torch.Tensor(weights).to(device=device)
+    for _ in range(dim):
+        weights_tensor = weights_tensor.unsqueeze(0)
+    for _ in range(total_dims - dim - 1):
+        weights_tensor = weights_tensor.unsqueeze(-1)
+    return weights_tensor
+
+def get_shape_for_dim(x_in: torch.Tensor, dim: int) -> list[int]:
+    total_dims = len(x_in.shape)
+    shape = []
+    for _ in range(dim):
+        shape.append(1)
+    shape.append(x_in.shape[dim])
+    for _ in range(total_dims - dim - 1):
+        shape.append(1)
+    return shape
+
+class ContextSchedules:
+    UNIFORM_LOOPED = "looped_uniform"
+    UNIFORM_STANDARD = "standard_uniform"
+    STATIC_STANDARD = "standard_static"
+    BATCHED = "batched"
+
+
+# from https://github.com/neggles/animatediff-cli/blob/main/src/animatediff/pipelines/context.py
+def create_windows_uniform_looped(num_frames: int, handler: IndexListContextHandler, model_options: dict[str]):
+    windows = []
+    if num_frames < handler.context_length:
+        windows.append(list(range(num_frames)))
+        return windows
+
+    context_stride = min(handler.context_stride, int(np.ceil(np.log2(num_frames / handler.context_length))) + 1)
+    # obtain uniform windows as normal, looping and all
+    for context_step in 1 << np.arange(context_stride):
+        pad = int(round(num_frames * ordered_halving(handler._step)))
+        for j in range(
+            int(ordered_halving(handler._step) * context_step) + pad,
+            num_frames + pad + (0 if handler.closed_loop else -handler.context_overlap),
+            (handler.context_length * context_step - handler.context_overlap),
+        ):
+            windows.append([e % num_frames for e in range(j, j + handler.context_length * context_step, context_step)])
+
+    return windows
+
+def create_windows_uniform_standard(num_frames: int, handler: IndexListContextHandler, model_options: dict[str]):
+    # unlike looped, uniform_straight does NOT allow windows that loop back to the beginning;
+    # instead, they get shifted to the corresponding end of the frames.
+    # in the case that a window (shifted or not) is identical to the previous one, it gets skipped.
+    windows = []
+    if num_frames <= handler.context_length:
+        windows.append(list(range(num_frames)))
+        return windows
+
+    context_stride = min(handler.context_stride, int(np.ceil(np.log2(num_frames / handler.context_length))) + 1)
+    # first, obtain uniform windows as normal, looping and all
+    for context_step in 1 << np.arange(context_stride):
+        pad = int(round(num_frames * ordered_halving(handler._step)))
+        for j in range(
+            int(ordered_halving(handler._step) * context_step) + pad,
+            num_frames + pad + (-handler.context_overlap),
+            (handler.context_length * context_step - handler.context_overlap),
+        ):
+            windows.append([e % num_frames for e in range(j, j + handler.context_length * context_step, context_step)])
+
+    # now that windows are created, shift any windows that loop, and delete duplicate windows
+    delete_idxs = []
+    win_i = 0
+    while win_i < len(windows):
+        # if window is rolls over itself, need to shift it
+        is_roll, roll_idx = does_window_roll_over(windows[win_i], num_frames)
+        if is_roll:
+            roll_val = windows[win_i][roll_idx]  # roll_val might not be 0 for windows of higher strides
+            shift_window_to_end(windows[win_i], num_frames=num_frames)
+            # check if next window (cyclical) is missing roll_val
+            if roll_val not in windows[(win_i+1) % len(windows)]:
+                # need to insert new window here - just insert window starting at roll_val
+                windows.insert(win_i+1, list(range(roll_val, roll_val + handler.context_length)))
+        # delete window if it's not unique
+        for pre_i in range(0, win_i):
+            if windows[win_i] == windows[pre_i]:
+                delete_idxs.append(win_i)
+                break
+        win_i += 1
+
+    # reverse delete_idxs so that they will be deleted in an order that doesn't break idx correlation
+    delete_idxs.reverse()
+    for i in delete_idxs:
+        windows.pop(i)
+
+    return windows
+
+
+def create_windows_static_standard(num_frames: int, handler: IndexListContextHandler, model_options: dict[str]):
+    windows = []
+    if num_frames <= handler.context_length:
+        windows.append(list(range(num_frames)))
+        return windows
+    # always return the same set of windows
+    delta = handler.context_length - handler.context_overlap
+    for start_idx in range(0, num_frames, delta):
+        # if past the end of frames, move start_idx back to allow same context_length
+        ending = start_idx + handler.context_length
+        if ending >= num_frames:
+            final_delta = ending - num_frames
+            final_start_idx = start_idx - final_delta
+            windows.append(list(range(final_start_idx, final_start_idx + handler.context_length)))
+            break
+        windows.append(list(range(start_idx, start_idx + handler.context_length)))
+    return windows
+
+
+def create_windows_batched(num_frames: int, handler: IndexListContextHandler, model_options: dict[str]):
+    windows = []
+    if num_frames <= handler.context_length:
+        windows.append(list(range(num_frames)))
+        return windows
+    # always return the same set of windows;
+    # no overlap, just cut up based on context_length;
+    # last window size will be different if num_frames % opts.context_length != 0
+    for start_idx in range(0, num_frames, handler.context_length):
+        windows.append(list(range(start_idx, min(start_idx + handler.context_length, num_frames))))
+    return windows
+
+
+def create_windows_default(num_frames: int, handler: IndexListContextHandler):
+    return [list(range(num_frames))]
+
+
+CONTEXT_MAPPING = {
+    ContextSchedules.UNIFORM_LOOPED: create_windows_uniform_looped,
+    ContextSchedules.UNIFORM_STANDARD: create_windows_uniform_standard,
+    ContextSchedules.STATIC_STANDARD: create_windows_static_standard,
+    ContextSchedules.BATCHED: create_windows_batched,
+}
+
+
+def get_matching_context_schedule(context_schedule: str) -> ContextSchedule:
+    func = CONTEXT_MAPPING.get(context_schedule, None)
+    if func is None:
+        raise ValueError(f"Unknown context_schedule '{context_schedule}'.")
+    return ContextSchedule(context_schedule, func)
+
+
+def get_context_weights(length: int, full_length: int, idxs: list[int], handler: IndexListContextHandler, sigma: torch.Tensor=None):
+    return handler.fuse_method.func(length, sigma=sigma, handler=handler, full_length=full_length, idxs=idxs)
+
+
+def create_weights_flat(length: int, **kwargs) -> list[float]:
+    # weight is the same for all
+    return [1.0] * length
+
+def create_weights_pyramid(length: int, **kwargs) -> list[float]:
+    # weight is based on the distance away from the edge of the context window;
+    # based on weighted average concept in FreeNoise paper
+    if length % 2 == 0:
+        max_weight = length // 2
+        weight_sequence = list(range(1, max_weight + 1, 1)) + list(range(max_weight, 0, -1))
+    else:
+        max_weight = (length + 1) // 2
+        weight_sequence = list(range(1, max_weight, 1)) + [max_weight] + list(range(max_weight - 1, 0, -1))
+    return weight_sequence
+
+def create_weights_overlap_linear(length: int, full_length: int, idxs: list[int], handler: IndexListContextHandler, **kwargs):
+    # based on code in Kijai's WanVideoWrapper: https://github.com/kijai/ComfyUI-WanVideoWrapper/blob/dbb2523b37e4ccdf45127e5ae33e31362f755c8e/nodes.py#L1302
+    # only expected overlap is given different weights
+    weights_torch = torch.ones((length))
+    # blend left-side on all except first window
+    if min(idxs) > 0:
+        ramp_up = torch.linspace(1e-37, 1, handler.context_overlap)
+        weights_torch[:handler.context_overlap] = ramp_up
+    # blend right-side on all except last window
+    if max(idxs) < full_length-1:
+        ramp_down = torch.linspace(1, 1e-37, handler.context_overlap)
+        weights_torch[-handler.context_overlap:] = ramp_down
+    return weights_torch
+
+class ContextFuseMethods:
+    FLAT = "flat"
+    PYRAMID = "pyramid"
+    RELATIVE = "relative"
+    OVERLAP_LINEAR = "overlap-linear"
+
+    LIST = [PYRAMID, FLAT, OVERLAP_LINEAR]
+    LIST_STATIC = [PYRAMID, RELATIVE, FLAT, OVERLAP_LINEAR]
+
+
+FUSE_MAPPING = {
+    ContextFuseMethods.FLAT: create_weights_flat,
+    ContextFuseMethods.PYRAMID: create_weights_pyramid,
+    ContextFuseMethods.RELATIVE: create_weights_pyramid,
+    ContextFuseMethods.OVERLAP_LINEAR: create_weights_overlap_linear,
+}
+
+def get_matching_fuse_method(fuse_method: str) -> ContextFuseMethod:
+    func = FUSE_MAPPING.get(fuse_method, None)
+    if func is None:
+        raise ValueError(f"Unknown fuse_method '{fuse_method}'.")
+    return ContextFuseMethod(fuse_method, func)
+
+# Returns fraction that has denominator that is a power of 2
+def ordered_halving(val):
+    # get binary value, padded with 0s for 64 bits
+    bin_str = f"{val:064b}"
+    # flip binary value, padding included
+    bin_flip = bin_str[::-1]
+    # convert binary to int
+    as_int = int(bin_flip, 2)
+    # divide by 1 << 64, equivalent to 2**64, or 18446744073709551616,
+    # or b10000000000000000000000000000000000000000000000000000000000000000 (1 with 64 zero's)
+    return as_int / (1 << 64)
+
+
+def get_missing_indexes(windows: list[list[int]], num_frames: int) -> list[int]:
+    all_indexes = list(range(num_frames))
+    for w in windows:
+        for val in w:
+            try:
+                all_indexes.remove(val)
+            except ValueError:
+                pass
+    return all_indexes
+
+
+def does_window_roll_over(window: list[int], num_frames: int) -> tuple[bool, int]:
+    prev_val = -1
+    for i, val in enumerate(window):
+        val = val % num_frames
+        if val < prev_val:
+            return True, i
+        prev_val = val
+    return False, -1
+
+
+def shift_window_to_start(window: list[int], num_frames: int):
+    start_val = window[0]
+    for i in range(len(window)):
+        # 1) subtract each element by start_val to move vals relative to the start of all frames
+        # 2) add num_frames and take modulus to get adjusted vals
+        window[i] = ((window[i] - start_val) + num_frames) % num_frames
+
+
+def shift_window_to_end(window: list[int], num_frames: int):
+    # 1) shift window to start
+    shift_window_to_start(window, num_frames)
+    end_val = window[-1]
+    end_delta = num_frames - end_val - 1
+    for i in range(len(window)):
+        # 2) add end_delta to each val to slide windows to end
+        window[i] = window[i] + end_delta

comfy/controlnet.py

@@ -36,6 +36,7 @@ import comfy.ldm.cascade.controlnet
 import comfy.cldm.mmdit
 import comfy.ldm.hydit.controlnet
 import comfy.ldm.flux.controlnet
+import comfy.ldm.qwen_image.controlnet
 import comfy.cldm.dit_embedder
 from typing import TYPE_CHECKING
 if TYPE_CHECKING:
@@ -236,11 +237,11 @@ class ControlNet(ControlBase):
             self.cond_hint = None
             compression_ratio = self.compression_ratio
             if self.vae is not None:
-                compression_ratio *= self.vae.downscale_ratio
+                compression_ratio *= self.vae.spacial_compression_encode()
             else:
                 if self.latent_format is not None:
                     raise ValueError("This Controlnet needs a VAE but none was provided, please use a ControlNetApply node with a VAE input and connect it.")
-            self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, x_noisy.shape[3] * compression_ratio, x_noisy.shape[2] * compression_ratio, self.upscale_algorithm, "center")
+            self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, x_noisy.shape[-1] * compression_ratio, x_noisy.shape[-2] * compression_ratio, self.upscale_algorithm, "center")
             self.cond_hint = self.preprocess_image(self.cond_hint)
             if self.vae is not None:
                 loaded_models = comfy.model_management.loaded_models(only_currently_used=True)
@@ -582,6 +583,15 @@ def load_controlnet_flux_instantx(sd, model_options={}):
     control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, concat_mask=concat_mask, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
     return control
 
+def load_controlnet_qwen_instantx(sd, model_options={}):
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(sd, model_options=model_options)
+    control_model = comfy.ldm.qwen_image.controlnet.QwenImageControlNetModel(operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
+    control_model = controlnet_load_state_dict(control_model, sd)
+    latent_format = comfy.latent_formats.Wan21()
+    extra_conds = []
+    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)
+    return control
+
 def convert_mistoline(sd):
     return comfy.utils.state_dict_prefix_replace(sd, {"single_controlnet_blocks.": "controlnet_single_blocks."})
 
@@ -655,8 +665,11 @@ def load_controlnet_state_dict(state_dict, model=None, model_options={}):
                 return load_controlnet_sd35(controlnet_data, model_options=model_options) #Stability sd3.5 format
             else:
                 return load_controlnet_mmdit(controlnet_data, model_options=model_options) #SD3 diffusers controlnet
+        elif "transformer_blocks.0.img_mlp.net.0.proj.weight" in controlnet_data:
+            return load_controlnet_qwen_instantx(controlnet_data, model_options=model_options)
         elif "controlnet_x_embedder.weight" in controlnet_data:
             return load_controlnet_flux_instantx(controlnet_data, model_options=model_options)
+
     elif "controlnet_blocks.0.linear.weight" in controlnet_data: #mistoline flux
         return load_controlnet_flux_xlabs_mistoline(convert_mistoline(controlnet_data), mistoline=True, model_options=model_options)
 

comfy/k_diffusion/sampling.py

@@ -853,6 +853,11 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
     return x
 
 
+@torch.no_grad()
+def sample_dpmpp_2m_sde_heun(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='heun'):
+    return sample_dpmpp_2m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
+
+
 @torch.no_grad()
 def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     """DPM-Solver++(3M) SDE."""
@@ -925,6 +930,16 @@ def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, di
     return sample_dpmpp_3m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler)
 
 
+@torch.no_grad()
+def sample_dpmpp_2m_sde_heun_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='heun'):
+    if len(sigmas) <= 1:
+        return x
+    extra_args = {} if extra_args is None else extra_args
+    sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
+    noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
+    return sample_dpmpp_2m_sde_heun(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
+
+
 @torch.no_grad()
 def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
     if len(sigmas) <= 1:

comfy/ldm/ace/attention.py

@@ -133,6 +133,7 @@ class Attention(nn.Module):
         hidden_states: torch.Tensor,
         encoder_hidden_states: Optional[torch.Tensor] = None,
         attention_mask: Optional[torch.Tensor] = None,
+        transformer_options={},
         **cross_attention_kwargs,
     ) -> torch.Tensor:
         return self.processor(
@@ -140,6 +141,7 @@ class Attention(nn.Module):
             hidden_states,
             encoder_hidden_states=encoder_hidden_states,
             attention_mask=attention_mask,
+            transformer_options=transformer_options,
             **cross_attention_kwargs,
         )
 
@@ -366,6 +368,7 @@ class CustomerAttnProcessor2_0:
         encoder_attention_mask: Optional[torch.FloatTensor] = None,
         rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
         rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        transformer_options={},
         *args,
         **kwargs,
     ) -> torch.Tensor:
@@ -433,7 +436,7 @@ class CustomerAttnProcessor2_0:
 
         # the output of sdp = (batch, num_heads, seq_len, head_dim)
         hidden_states = optimized_attention(
-            query, key, value, heads=query.shape[1], mask=attention_mask, skip_reshape=True,
+            query, key, value, heads=query.shape[1], mask=attention_mask, skip_reshape=True, transformer_options=transformer_options,
         ).to(query.dtype)
 
         # linear proj
@@ -697,6 +700,7 @@ class LinearTransformerBlock(nn.Module):
         rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
         rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
         temb: torch.FloatTensor = None,
+        transformer_options={},
     ):
 
         N = hidden_states.shape[0]
@@ -720,6 +724,7 @@ class LinearTransformerBlock(nn.Module):
                 encoder_attention_mask=encoder_attention_mask,
                 rotary_freqs_cis=rotary_freqs_cis,
                 rotary_freqs_cis_cross=rotary_freqs_cis_cross,
+                transformer_options=transformer_options,
             )
         else:
             attn_output, _ = self.attn(
@@ -729,6 +734,7 @@ class LinearTransformerBlock(nn.Module):
                 encoder_attention_mask=None,
                 rotary_freqs_cis=rotary_freqs_cis,
                 rotary_freqs_cis_cross=None,
+                transformer_options=transformer_options,
             )
 
         if self.use_adaln_single:
@@ -743,6 +749,7 @@ class LinearTransformerBlock(nn.Module):
                 encoder_attention_mask=encoder_attention_mask,
                 rotary_freqs_cis=rotary_freqs_cis,
                 rotary_freqs_cis_cross=rotary_freqs_cis_cross,
+                transformer_options=transformer_options,
             )
             hidden_states = attn_output + hidden_states
 

comfy/ldm/ace/model.py

@@ -19,6 +19,7 @@ import torch
 from torch import nn
 
 import comfy.model_management
+import comfy.patcher_extension
 
 from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps
 from .attention import LinearTransformerBlock, t2i_modulate
@@ -313,6 +314,7 @@ class ACEStepTransformer2DModel(nn.Module):
         output_length: int = 0,
         block_controlnet_hidden_states: Optional[Union[List[torch.Tensor], torch.Tensor]] = None,
         controlnet_scale: Union[float, torch.Tensor] = 1.0,
+        transformer_options={},
     ):
         embedded_timestep = self.timestep_embedder(self.time_proj(timestep).to(dtype=hidden_states.dtype))
         temb = self.t_block(embedded_timestep)
@@ -338,12 +340,34 @@ class ACEStepTransformer2DModel(nn.Module):
                 rotary_freqs_cis=rotary_freqs_cis,
                 rotary_freqs_cis_cross=encoder_rotary_freqs_cis,
                 temb=temb,
+                transformer_options=transformer_options,
             )
 
         output = self.final_layer(hidden_states, embedded_timestep, output_length)
         return output
 
-    def forward(
+    def forward(self,
+        x,
+        timestep,
+        attention_mask=None,
+        context: Optional[torch.Tensor] = None,
+        text_attention_mask: Optional[torch.LongTensor] = None,
+        speaker_embeds: Optional[torch.FloatTensor] = None,
+        lyric_token_idx: Optional[torch.LongTensor] = None,
+        lyric_mask: Optional[torch.LongTensor] = None,
+        block_controlnet_hidden_states: Optional[Union[List[torch.Tensor], torch.Tensor]] = None,
+        controlnet_scale: Union[float, torch.Tensor] = 1.0,
+        lyrics_strength=1.0,
+        **kwargs
+    ):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, kwargs.get("transformer_options", {}))
+        ).execute(x, timestep, attention_mask, context, text_attention_mask, speaker_embeds, lyric_token_idx, lyric_mask, block_controlnet_hidden_states,
+                  controlnet_scale, lyrics_strength, **kwargs)
+
+    def _forward(
         self,
         x,
         timestep,
@@ -371,6 +395,7 @@ class ACEStepTransformer2DModel(nn.Module):
 
         output_length = hidden_states.shape[-1]
 
+        transformer_options = kwargs.get("transformer_options", {})
         output = self.decode(
             hidden_states=hidden_states,
             attention_mask=attention_mask,
@@ -380,6 +405,7 @@ class ACEStepTransformer2DModel(nn.Module):
             output_length=output_length,
             block_controlnet_hidden_states=block_controlnet_hidden_states,
             controlnet_scale=controlnet_scale,
+            transformer_options=transformer_options,
         )
 
         return output

comfy/ldm/audio/dit.py

@@ -298,7 +298,8 @@ class Attention(nn.Module):
         mask = None,
         context_mask = None,
         rotary_pos_emb = None,
-        causal = None
+        causal = None,
+        transformer_options={},
     ):
         h, kv_h, has_context = self.num_heads, self.kv_heads, context is not None
 
@@ -363,7 +364,7 @@ class Attention(nn.Module):
             heads_per_kv_head = h // kv_h
             k, v = map(lambda t: t.repeat_interleave(heads_per_kv_head, dim = 1), (k, v))
 
-        out = optimized_attention(q, k, v, h, skip_reshape=True)
+        out = optimized_attention(q, k, v, h, skip_reshape=True, transformer_options=transformer_options)
         out = self.to_out(out)
 
         if mask is not None:
@@ -488,7 +489,8 @@ class TransformerBlock(nn.Module):
         global_cond=None,
         mask = None,
         context_mask = None,
-        rotary_pos_emb = None
+        rotary_pos_emb = None,
+        transformer_options={}
     ):
         if self.global_cond_dim is not None and self.global_cond_dim > 0 and global_cond is not None:
 
@@ -498,12 +500,12 @@ class TransformerBlock(nn.Module):
             residual = x
             x = self.pre_norm(x)
             x = x * (1 + scale_self) + shift_self
-            x = self.self_attn(x, mask = mask, rotary_pos_emb = rotary_pos_emb)
+            x = self.self_attn(x, mask = mask, rotary_pos_emb = rotary_pos_emb, transformer_options=transformer_options)
             x = x * torch.sigmoid(1 - gate_self)
             x = x + residual
 
             if context is not None:
-                x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask)
+                x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask, transformer_options=transformer_options)
 
             if self.conformer is not None:
                 x = x + self.conformer(x)
@@ -517,10 +519,10 @@ class TransformerBlock(nn.Module):
             x = x + residual
 
         else:
-            x = x + self.self_attn(self.pre_norm(x), mask = mask, rotary_pos_emb = rotary_pos_emb)
+            x = x + self.self_attn(self.pre_norm(x), mask = mask, rotary_pos_emb = rotary_pos_emb, transformer_options=transformer_options)
 
             if context is not None:
-                x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask)
+                x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask, transformer_options=transformer_options)
 
             if self.conformer is not None:
                 x = x + self.conformer(x)
@@ -606,7 +608,8 @@ class ContinuousTransformer(nn.Module):
         return_info = False,
         **kwargs
     ):
-        patches_replace = kwargs.get("transformer_options", {}).get("patches_replace", {})
+        transformer_options = kwargs.get("transformer_options", {})
+        patches_replace = transformer_options.get("patches_replace", {})
         batch, seq, device = *x.shape[:2], x.device
         context = kwargs["context"]
 
@@ -645,13 +648,13 @@ class ContinuousTransformer(nn.Module):
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"] = layer(args["img"], rotary_pos_emb=args["pe"], global_cond=args["vec"], context=args["txt"])
+                    out["img"] = layer(args["img"], rotary_pos_emb=args["pe"], global_cond=args["vec"], context=args["txt"], transformer_options=args["transformer_options"])
                     return out
 
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": global_cond, "pe": rotary_pos_emb}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": global_cond, "pe": rotary_pos_emb, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 x = out["img"]
             else:
-                x = layer(x, rotary_pos_emb = rotary_pos_emb, global_cond=global_cond, context=context)
+                x = layer(x, rotary_pos_emb = rotary_pos_emb, global_cond=global_cond, context=context, transformer_options=transformer_options)
             # x = checkpoint(layer, x, rotary_pos_emb = rotary_pos_emb, global_cond=global_cond, **kwargs)
 
             if return_info:

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.patcher_extension
 import comfy.ldm.common_dit
 
 def modulate(x, shift, scale):
@@ -84,7 +85,7 @@ class SingleAttention(nn.Module):
         )
 
     #@torch.compile()
-    def forward(self, c):
+    def forward(self, c, transformer_options={}):
 
         bsz, seqlen1, _ = c.shape
 
@@ -94,7 +95,7 @@ class SingleAttention(nn.Module):
         v = v.view(bsz, seqlen1, self.n_heads, self.head_dim)
         q, k = self.q_norm1(q), self.k_norm1(k)
 
-        output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True)
+        output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True, transformer_options=transformer_options)
         c = self.w1o(output)
         return c
 
@@ -143,7 +144,7 @@ class DoubleAttention(nn.Module):
 
 
     #@torch.compile()
-    def forward(self, c, x):
+    def forward(self, c, x, transformer_options={}):
 
         bsz, seqlen1, _ = c.shape
         bsz, seqlen2, _ = x.shape
@@ -167,7 +168,7 @@ class DoubleAttention(nn.Module):
             torch.cat([cv, xv], dim=1),
         )
 
-        output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True)
+        output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True, transformer_options=transformer_options)
 
         c, x = output.split([seqlen1, seqlen2], dim=1)
         c = self.w1o(c)
@@ -206,7 +207,7 @@ class MMDiTBlock(nn.Module):
         self.is_last = is_last
 
     #@torch.compile()
-    def forward(self, c, x, global_cond, **kwargs):
+    def forward(self, c, x, global_cond, transformer_options={}, **kwargs):
 
         cres, xres = c, x
 
@@ -224,7 +225,7 @@ class MMDiTBlock(nn.Module):
         x = modulate(self.normX1(x), xshift_msa, xscale_msa)
 
         # attention
-        c, x = self.attn(c, x)
+        c, x = self.attn(c, x, transformer_options=transformer_options)
 
 
         c = self.normC2(cres + cgate_msa.unsqueeze(1) * c)
@@ -254,13 +255,13 @@ class DiTBlock(nn.Module):
         self.mlp = MLP(dim, hidden_dim=dim * 4, dtype=dtype, device=device, operations=operations)
 
     #@torch.compile()
-    def forward(self, cx, global_cond, **kwargs):
+    def forward(self, cx, global_cond, transformer_options={}, **kwargs):
         cxres = cx
         shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.modCX(
             global_cond
         ).chunk(6, dim=1)
         cx = modulate(self.norm1(cx), shift_msa, scale_msa)
-        cx = self.attn(cx)
+        cx = self.attn(cx, transformer_options=transformer_options)
         cx = self.norm2(cxres + gate_msa.unsqueeze(1) * cx)
         mlpout = self.mlp(modulate(cx, shift_mlp, scale_mlp))
         cx = gate_mlp.unsqueeze(1) * mlpout
@@ -436,6 +437,13 @@ class MMDiT(nn.Module):
         return x + pos_encoding.reshape(1, -1, self.positional_encoding.shape[-1])
 
     def forward(self, x, timestep, context, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, transformer_options={}, **kwargs):
         patches_replace = transformer_options.get("patches_replace", {})
         # patchify x, add PE
         b, c, h, w = x.shape
@@ -465,13 +473,14 @@ class MMDiT(nn.Module):
                         out = {}
                         out["txt"], out["img"] = layer(args["txt"],
                                                        args["img"],
-                                                       args["vec"])
+                                                       args["vec"],
+                                                       transformer_options=args["transformer_options"])
                         return out
-                    out = blocks_replace[("double_block", i)]({"img": x, "txt": c, "vec": global_cond}, {"original_block": block_wrap})
+                    out = blocks_replace[("double_block", i)]({"img": x, "txt": c, "vec": global_cond, "transformer_options": transformer_options}, {"original_block": block_wrap})
                     c = out["txt"]
                     x = out["img"]
                 else:
-                    c, x = layer(c, x, global_cond, **kwargs)
+                    c, x = layer(c, x, global_cond, transformer_options=transformer_options, **kwargs)
 
         if len(self.single_layers) > 0:
             c_len = c.size(1)
@@ -480,13 +489,13 @@ class MMDiT(nn.Module):
                 if ("single_block", i) in blocks_replace:
                     def block_wrap(args):
                         out = {}
-                        out["img"] = layer(args["img"], args["vec"])
+                        out["img"] = layer(args["img"], args["vec"], transformer_options=args["transformer_options"])
                         return out
 
-                    out = blocks_replace[("single_block", i)]({"img": cx, "vec": global_cond}, {"original_block": block_wrap})
+                    out = blocks_replace[("single_block", i)]({"img": cx, "vec": global_cond, "transformer_options": transformer_options}, {"original_block": block_wrap})
                     cx = out["img"]
                 else:
-                    cx = layer(cx, global_cond, **kwargs)
+                    cx = layer(cx, global_cond, transformer_options=transformer_options, **kwargs)
 
             x = cx[:, c_len:]
 

comfy/ldm/cascade/common.py

@@ -32,12 +32,12 @@ class OptimizedAttention(nn.Module):
 
         self.out_proj = operations.Linear(c, c, bias=True, dtype=dtype, device=device)
 
-    def forward(self, q, k, v):
+    def forward(self, q, k, v, transformer_options={}):
         q = self.to_q(q)
         k = self.to_k(k)
         v = self.to_v(v)
 
-        out = optimized_attention(q, k, v, self.heads)
+        out = optimized_attention(q, k, v, self.heads, transformer_options=transformer_options)
 
         return self.out_proj(out)
 
@@ -47,13 +47,13 @@ class Attention2D(nn.Module):
         self.attn = OptimizedAttention(c, nhead, dtype=dtype, device=device, operations=operations)
         # self.attn = nn.MultiheadAttention(c, nhead, dropout=dropout, bias=True, batch_first=True, dtype=dtype, device=device)
 
-    def forward(self, x, kv, self_attn=False):
+    def forward(self, x, kv, self_attn=False, transformer_options={}):
         orig_shape = x.shape
         x = x.view(x.size(0), x.size(1), -1).permute(0, 2, 1)  # Bx4xHxW -> Bx(HxW)x4
         if self_attn:
             kv = torch.cat([x, kv], dim=1)
         # x = self.attn(x, kv, kv, need_weights=False)[0]
-        x = self.attn(x, kv, kv)
+        x = self.attn(x, kv, kv, transformer_options=transformer_options)
         x = x.permute(0, 2, 1).view(*orig_shape)
         return x
 
@@ -114,9 +114,9 @@ class AttnBlock(nn.Module):
             operations.Linear(c_cond, c, dtype=dtype, device=device)
         )
 
-    def forward(self, x, kv):
+    def forward(self, x, kv, transformer_options={}):
         kv = self.kv_mapper(kv)
-        x = x + self.attention(self.norm(x), kv, self_attn=self.self_attn)
+        x = x + self.attention(self.norm(x), kv, self_attn=self.self_attn, transformer_options=transformer_options)
         return x
 
 

comfy/ldm/cascade/stage_b.py

@@ -173,7 +173,7 @@ class StageB(nn.Module):
         clip = self.clip_norm(clip)
         return clip
 
-    def _down_encode(self, x, r_embed, clip):
+    def _down_encode(self, x, r_embed, clip, transformer_options={}):
         level_outputs = []
         block_group = zip(self.down_blocks, self.down_downscalers, self.down_repeat_mappers)
         for down_block, downscaler, repmap in block_group:
@@ -187,7 +187,7 @@ class StageB(nn.Module):
                     elif isinstance(block, AttnBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   AttnBlock)):
-                        x = block(x, clip)
+                        x = block(x, clip, transformer_options=transformer_options)
                     elif isinstance(block, TimestepBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   TimestepBlock)):
@@ -199,7 +199,7 @@ class StageB(nn.Module):
             level_outputs.insert(0, x)
         return level_outputs
 
-    def _up_decode(self, level_outputs, r_embed, clip):
+    def _up_decode(self, level_outputs, r_embed, clip, transformer_options={}):
         x = level_outputs[0]
         block_group = zip(self.up_blocks, self.up_upscalers, self.up_repeat_mappers)
         for i, (up_block, upscaler, repmap) in enumerate(block_group):
@@ -216,7 +216,7 @@ class StageB(nn.Module):
                     elif isinstance(block, AttnBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   AttnBlock)):
-                        x = block(x, clip)
+                        x = block(x, clip, transformer_options=transformer_options)
                     elif isinstance(block, TimestepBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   TimestepBlock)):
@@ -228,7 +228,7 @@ class StageB(nn.Module):
             x = upscaler(x)
         return x
 
-    def forward(self, x, r, effnet, clip, pixels=None, **kwargs):
+    def forward(self, x, r, effnet, clip, pixels=None, transformer_options={}, **kwargs):
         if pixels is None:
             pixels = x.new_zeros(x.size(0), 3, 8, 8)
 
@@ -245,8 +245,8 @@ class StageB(nn.Module):
             nn.functional.interpolate(effnet, size=x.shape[-2:], mode='bilinear', align_corners=True))
         x = x + nn.functional.interpolate(self.pixels_mapper(pixels), size=x.shape[-2:], mode='bilinear',
                                           align_corners=True)
-        level_outputs = self._down_encode(x, r_embed, clip)
-        x = self._up_decode(level_outputs, r_embed, clip)
+        level_outputs = self._down_encode(x, r_embed, clip, transformer_options=transformer_options)
+        x = self._up_decode(level_outputs, r_embed, clip, transformer_options=transformer_options)
         return self.clf(x)
 
     def update_weights_ema(self, src_model, beta=0.999):

comfy/ldm/cascade/stage_c.py

@@ -182,7 +182,7 @@ class StageC(nn.Module):
         clip = self.clip_norm(clip)
         return clip
 
-    def _down_encode(self, x, r_embed, clip, cnet=None):
+    def _down_encode(self, x, r_embed, clip, cnet=None, transformer_options={}):
         level_outputs = []
         block_group = zip(self.down_blocks, self.down_downscalers, self.down_repeat_mappers)
         for down_block, downscaler, repmap in block_group:
@@ -201,7 +201,7 @@ class StageC(nn.Module):
                     elif isinstance(block, AttnBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   AttnBlock)):
-                        x = block(x, clip)
+                        x = block(x, clip, transformer_options=transformer_options)
                     elif isinstance(block, TimestepBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   TimestepBlock)):
@@ -213,7 +213,7 @@ class StageC(nn.Module):
             level_outputs.insert(0, x)
         return level_outputs
 
-    def _up_decode(self, level_outputs, r_embed, clip, cnet=None):
+    def _up_decode(self, level_outputs, r_embed, clip, cnet=None, transformer_options={}):
         x = level_outputs[0]
         block_group = zip(self.up_blocks, self.up_upscalers, self.up_repeat_mappers)
         for i, (up_block, upscaler, repmap) in enumerate(block_group):
@@ -235,7 +235,7 @@ class StageC(nn.Module):
                     elif isinstance(block, AttnBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   AttnBlock)):
-                        x = block(x, clip)
+                        x = block(x, clip, transformer_options=transformer_options)
                     elif isinstance(block, TimestepBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   TimestepBlock)):
@@ -247,7 +247,7 @@ class StageC(nn.Module):
             x = upscaler(x)
         return x
 
-    def forward(self, x, r, clip_text, clip_text_pooled, clip_img, control=None, **kwargs):
+    def forward(self, x, r, clip_text, clip_text_pooled, clip_img, control=None, transformer_options={}, **kwargs):
         # Process the conditioning embeddings
         r_embed = self.gen_r_embedding(r).to(dtype=x.dtype)
         for c in self.t_conds:
@@ -262,8 +262,8 @@ class StageC(nn.Module):
 
         # Model Blocks
         x = self.embedding(x)
-        level_outputs = self._down_encode(x, r_embed, clip, cnet)
-        x = self._up_decode(level_outputs, r_embed, clip, cnet)
+        level_outputs = self._down_encode(x, r_embed, clip, cnet, transformer_options=transformer_options)
+        x = self._up_decode(level_outputs, r_embed, clip, cnet, transformer_options=transformer_options)
         return self.clf(x)
 
     def update_weights_ema(self, src_model, beta=0.999):

comfy/ldm/chroma/layers.py

@@ -76,7 +76,7 @@ class DoubleStreamBlock(nn.Module):
         )
         self.flipped_img_txt = flipped_img_txt
 
-    def forward(self, img: Tensor, txt: Tensor, pe: Tensor, vec: Tensor, attn_mask=None):
+    def forward(self, img: Tensor, txt: Tensor, pe: Tensor, vec: Tensor, attn_mask=None, transformer_options={}):
         (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
 
         # prepare image for attention
@@ -95,7 +95,7 @@ class DoubleStreamBlock(nn.Module):
         attn = attention(torch.cat((txt_q, img_q), dim=2),
                          torch.cat((txt_k, img_k), dim=2),
                          torch.cat((txt_v, img_v), dim=2),
-                         pe=pe, mask=attn_mask)
+                         pe=pe, mask=attn_mask, transformer_options=transformer_options)
 
         txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
 
@@ -148,7 +148,7 @@ class SingleStreamBlock(nn.Module):
 
         self.mlp_act = nn.GELU(approximate="tanh")
 
-    def forward(self, x: Tensor, pe: Tensor, vec: Tensor, attn_mask=None) -> Tensor:
+    def forward(self, x: Tensor, pe: Tensor, vec: Tensor, attn_mask=None, transformer_options={}) -> Tensor:
         mod = vec
         x_mod = torch.addcmul(mod.shift, 1 + mod.scale, self.pre_norm(x))
         qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
@@ -157,7 +157,7 @@ class SingleStreamBlock(nn.Module):
         q, k = self.norm(q, k, v)
 
         # compute attention
-        attn = attention(q, k, v, pe=pe, mask=attn_mask)
+        attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
         # compute activation in mlp stream, cat again and run second linear layer
         output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
         x.addcmul_(mod.gate, output)

comfy/ldm/chroma/model.py

@@ -5,6 +5,7 @@ from dataclasses import dataclass
 import torch
 from torch import Tensor, nn
 from einops import rearrange, repeat
+import comfy.patcher_extension
 import comfy.ldm.common_dit
 
 from comfy.ldm.flux.layers import (
@@ -192,14 +193,16 @@ class Chroma(nn.Module):
                                                        txt=args["txt"],
                                                        vec=args["vec"],
                                                        pe=args["pe"],
-                                                       attn_mask=args.get("attn_mask"))
+                                                       attn_mask=args.get("attn_mask"),
+                                                       transformer_options=args.get("transformer_options"))
                         return out
 
                     out = blocks_replace[("double_block", i)]({"img": img,
                                                                "txt": txt,
                                                                "vec": double_mod,
                                                                "pe": pe,
-                                                               "attn_mask": attn_mask},
+                                                               "attn_mask": attn_mask,
+                                                               "transformer_options": transformer_options},
                                                               {"original_block": block_wrap})
                     txt = out["txt"]
                     img = out["img"]
@@ -208,7 +211,8 @@ class Chroma(nn.Module):
                                      txt=txt,
                                      vec=double_mod,
                                      pe=pe,
-                                     attn_mask=attn_mask)
+                                     attn_mask=attn_mask,
+                                     transformer_options=transformer_options)
 
                 if control is not None: # Controlnet
                     control_i = control.get("input")
@@ -228,17 +232,19 @@ class Chroma(nn.Module):
                         out["img"] = block(args["img"],
                                            vec=args["vec"],
                                            pe=args["pe"],
-                                           attn_mask=args.get("attn_mask"))
+                                           attn_mask=args.get("attn_mask"),
+                                           transformer_options=args.get("transformer_options"))
                         return out
 
                     out = blocks_replace[("single_block", i)]({"img": img,
                                                                "vec": single_mod,
                                                                "pe": pe,
-                                                               "attn_mask": attn_mask},
+                                                               "attn_mask": attn_mask,
+                                                               "transformer_options": transformer_options},
                                                               {"original_block": block_wrap})
                     img = out["img"]
                 else:
-                    img = block(img, vec=single_mod, pe=pe, attn_mask=attn_mask)
+                    img = block(img, vec=single_mod, pe=pe, attn_mask=attn_mask, transformer_options=transformer_options)
 
                 if control is not None: # Controlnet
                     control_o = control.get("output")
@@ -253,6 +259,13 @@ class Chroma(nn.Module):
         return img
 
     def forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, guidance, control, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs):
         bs, c, h, w = x.shape
         x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
 

comfy/ldm/cosmos/blocks.py

@@ -176,6 +176,7 @@ class Attention(nn.Module):
         context=None,
         mask=None,
         rope_emb=None,
+        transformer_options={},
         **kwargs,
     ):
         """
@@ -184,7 +185,7 @@ class Attention(nn.Module):
             context (Optional[Tensor]): The key tensor of shape [B, Mk, K] or use x as context [self attention] if None
         """
         q, k, v = self.cal_qkv(x, context, mask, rope_emb=rope_emb, **kwargs)
-        out = optimized_attention(q, k, v, self.heads, skip_reshape=True, mask=mask, skip_output_reshape=True)
+        out = optimized_attention(q, k, v, self.heads, skip_reshape=True, mask=mask, skip_output_reshape=True, transformer_options=transformer_options)
         del q, k, v
         out = rearrange(out, " b n s c -> s b (n c)")
         return self.to_out(out)
@@ -546,6 +547,7 @@ class VideoAttn(nn.Module):
         context: Optional[torch.Tensor] = None,
         crossattn_mask: Optional[torch.Tensor] = None,
         rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
+        transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
         """
         Forward pass for video attention.
@@ -571,6 +573,7 @@ class VideoAttn(nn.Module):
             context_M_B_D,
             crossattn_mask,
             rope_emb=rope_emb_L_1_1_D,
+            transformer_options=transformer_options,
         )
         x_T_H_W_B_D = rearrange(x_THW_B_D, "(t h w) b d -> t h w b d", h=H, w=W)
         return x_T_H_W_B_D
@@ -665,6 +668,7 @@ class DITBuildingBlock(nn.Module):
         crossattn_mask: Optional[torch.Tensor] = None,
         rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
         adaln_lora_B_3D: Optional[torch.Tensor] = None,
+        transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
         """
         Forward pass for dynamically configured blocks with adaptive normalization.
@@ -702,6 +706,7 @@ class DITBuildingBlock(nn.Module):
                 adaln_norm_state(self.norm_state, x, scale_1_1_1_B_D, shift_1_1_1_B_D),
                 context=None,
                 rope_emb_L_1_1_D=rope_emb_L_1_1_D,
+                transformer_options=transformer_options,
             )
         elif self.block_type in ["cross_attn", "ca"]:
             x = x + gate_1_1_1_B_D * self.block(
@@ -709,6 +714,7 @@ class DITBuildingBlock(nn.Module):
                 context=crossattn_emb,
                 crossattn_mask=crossattn_mask,
                 rope_emb_L_1_1_D=rope_emb_L_1_1_D,
+                transformer_options=transformer_options,
             )
         else:
             raise ValueError(f"Unknown block type: {self.block_type}")
@@ -784,6 +790,7 @@ class GeneralDITTransformerBlock(nn.Module):
         crossattn_mask: Optional[torch.Tensor] = None,
         rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
         adaln_lora_B_3D: Optional[torch.Tensor] = None,
+        transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
         for block in self.blocks:
             x = block(
@@ -793,5 +800,6 @@ class GeneralDITTransformerBlock(nn.Module):
                 crossattn_mask,
                 rope_emb_L_1_1_D=rope_emb_L_1_1_D,
                 adaln_lora_B_3D=adaln_lora_B_3D,
+                transformer_options=transformer_options,
             )
         return x

comfy/ldm/cosmos/model.py

@@ -27,6 +27,8 @@ from torchvision import transforms
 from enum import Enum
 import logging
 
+import comfy.patcher_extension
+
 from .blocks import (
     FinalLayer,
     GeneralDITTransformerBlock,
@@ -435,6 +437,42 @@ class GeneralDIT(nn.Module):
         latent_condition_sigma: Optional[torch.Tensor] = None,
         condition_video_augment_sigma: Optional[torch.Tensor] = None,
         **kwargs,
+    ):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, kwargs.get("transformer_options", {}))
+        ).execute(x,
+                timesteps,
+                context,
+                attention_mask,
+                fps,
+                image_size,
+                padding_mask,
+                scalar_feature,
+                data_type,
+                latent_condition,
+                latent_condition_sigma,
+                condition_video_augment_sigma,
+                **kwargs)
+
+    def _forward(
+        self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        context: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        # crossattn_emb: torch.Tensor,
+        # crossattn_mask: Optional[torch.Tensor] = None,
+        fps: Optional[torch.Tensor] = None,
+        image_size: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        scalar_feature: Optional[torch.Tensor] = None,
+        data_type: Optional[DataType] = DataType.VIDEO,
+        latent_condition: Optional[torch.Tensor] = None,
+        latent_condition_sigma: Optional[torch.Tensor] = None,
+        condition_video_augment_sigma: Optional[torch.Tensor] = None,
+        **kwargs,
     ):
         """
         Args:
@@ -482,6 +520,7 @@ class GeneralDIT(nn.Module):
                 x.shape == extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D.shape
             ), f"{x.shape} != {extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D.shape} {original_shape}"
 
+        transformer_options = kwargs.get("transformer_options", {})
         for _, block in self.blocks.items():
             assert (
                 self.blocks["block0"].x_format == block.x_format
@@ -496,6 +535,7 @@ class GeneralDIT(nn.Module):
                 crossattn_mask,
                 rope_emb_L_1_1_D=rope_emb_L_1_1_D,
                 adaln_lora_B_3D=adaln_lora_B_3D,
+                transformer_options=transformer_options,
             )
 
         x_B_T_H_W_D = rearrange(x, "T H W B D -> B T H W D")

comfy/ldm/cosmos/predict2.py

@@ -11,6 +11,7 @@ import math
 from .position_embedding import VideoRopePosition3DEmb, LearnablePosEmbAxis
 from torchvision import transforms
 
+import comfy.patcher_extension
 from comfy.ldm.modules.attention import optimized_attention
 
 def apply_rotary_pos_emb(
@@ -43,7 +44,7 @@ class GPT2FeedForward(nn.Module):
         return x
 
 
-def torch_attention_op(q_B_S_H_D: torch.Tensor, k_B_S_H_D: torch.Tensor, v_B_S_H_D: torch.Tensor) -> torch.Tensor:
+def torch_attention_op(q_B_S_H_D: torch.Tensor, k_B_S_H_D: torch.Tensor, v_B_S_H_D: torch.Tensor, transformer_options: Optional[dict] = {}) -> torch.Tensor:
     """Computes multi-head attention using PyTorch's native implementation.
 
     This function provides a PyTorch backend alternative to Transformer Engine's attention operation.
@@ -70,7 +71,7 @@ def torch_attention_op(q_B_S_H_D: torch.Tensor, k_B_S_H_D: torch.Tensor, v_B_S_H
     q_B_H_S_D = rearrange(q_B_S_H_D, "b ... h k -> b h ... k").view(in_q_shape[0], in_q_shape[-2], -1, in_q_shape[-1])
     k_B_H_S_D = rearrange(k_B_S_H_D, "b ... h v -> b h ... v").view(in_k_shape[0], in_k_shape[-2], -1, in_k_shape[-1])
     v_B_H_S_D = rearrange(v_B_S_H_D, "b ... h v -> b h ... v").view(in_k_shape[0], in_k_shape[-2], -1, in_k_shape[-1])
-    return optimized_attention(q_B_H_S_D, k_B_H_S_D, v_B_H_S_D, in_q_shape[-2], skip_reshape=True)
+    return optimized_attention(q_B_H_S_D, k_B_H_S_D, v_B_H_S_D, in_q_shape[-2], skip_reshape=True, transformer_options=transformer_options)
 
 
 class Attention(nn.Module):
@@ -179,8 +180,8 @@ class Attention(nn.Module):
 
         return q, k, v
 
-    def compute_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> torch.Tensor:
-        result = self.attn_op(q, k, v)  # [B, S, H, D]
+    def compute_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, transformer_options: Optional[dict] = {}) -> torch.Tensor:
+        result = self.attn_op(q, k, v, transformer_options=transformer_options)  # [B, S, H, D]
         return self.output_dropout(self.output_proj(result))
 
     def forward(
@@ -188,6 +189,7 @@ class Attention(nn.Module):
         x: torch.Tensor,
         context: Optional[torch.Tensor] = None,
         rope_emb: Optional[torch.Tensor] = None,
+        transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
         """
         Args:
@@ -195,7 +197,7 @@ class Attention(nn.Module):
             context (Optional[Tensor]): The key tensor of shape [B, Mk, K] or use x as context [self attention] if None
         """
         q, k, v = self.compute_qkv(x, context, rope_emb=rope_emb)
-        return self.compute_attention(q, k, v)
+        return self.compute_attention(q, k, v, transformer_options=transformer_options)
 
 
 class Timesteps(nn.Module):
@@ -458,6 +460,7 @@ class Block(nn.Module):
         rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
         adaln_lora_B_T_3D: Optional[torch.Tensor] = None,
         extra_per_block_pos_emb: Optional[torch.Tensor] = None,
+        transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
         if extra_per_block_pos_emb is not None:
             x_B_T_H_W_D = x_B_T_H_W_D + extra_per_block_pos_emb
@@ -511,6 +514,7 @@ class Block(nn.Module):
                 rearrange(normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
                 None,
                 rope_emb=rope_emb_L_1_1_D,
+                transformer_options=transformer_options,
             ),
             "b (t h w) d -> b t h w d",
             t=T,
@@ -524,6 +528,7 @@ class Block(nn.Module):
             layer_norm_cross_attn: Callable,
             _scale_cross_attn_B_T_1_1_D: torch.Tensor,
             _shift_cross_attn_B_T_1_1_D: torch.Tensor,
+            transformer_options: Optional[dict] = {},
         ) -> torch.Tensor:
             _normalized_x_B_T_H_W_D = _fn(
                 _x_B_T_H_W_D, layer_norm_cross_attn, _scale_cross_attn_B_T_1_1_D, _shift_cross_attn_B_T_1_1_D
@@ -533,6 +538,7 @@ class Block(nn.Module):
                     rearrange(_normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
                     crossattn_emb,
                     rope_emb=rope_emb_L_1_1_D,
+                    transformer_options=transformer_options,
                 ),
                 "b (t h w) d -> b t h w d",
                 t=T,
@@ -546,6 +552,7 @@ class Block(nn.Module):
             self.layer_norm_cross_attn,
             scale_cross_attn_B_T_1_1_D,
             shift_cross_attn_B_T_1_1_D,
+            transformer_options=transformer_options,
         )
         x_B_T_H_W_D = result_B_T_H_W_D * gate_cross_attn_B_T_1_1_D + x_B_T_H_W_D
 
@@ -805,7 +812,21 @@ class MiniTrainDIT(nn.Module):
         )
         return x_B_C_Tt_Hp_Wp
 
-    def forward(
+    def forward(self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        context: torch.Tensor,
+        fps: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, kwargs.get("transformer_options", {}))
+        ).execute(x, timesteps, context, fps, padding_mask, **kwargs)
+
+    def _forward(
         self,
         x: torch.Tensor,
         timesteps: torch.Tensor,
@@ -850,6 +871,7 @@ class MiniTrainDIT(nn.Module):
             "rope_emb_L_1_1_D": rope_emb_L_1_1_D.unsqueeze(1).unsqueeze(0),
             "adaln_lora_B_T_3D": adaln_lora_B_T_3D,
             "extra_per_block_pos_emb": extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D,
+            "transformer_options": kwargs.get("transformer_options", {}),
         }
         for block in self.blocks:
             x_B_T_H_W_D = block(

comfy/ldm/flux/layers.py

@@ -159,7 +159,7 @@ class DoubleStreamBlock(nn.Module):
         )
         self.flipped_img_txt = flipped_img_txt
 
-    def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None):
+    def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}):
         img_mod1, img_mod2 = self.img_mod(vec)
         txt_mod1, txt_mod2 = self.txt_mod(vec)
 
@@ -182,7 +182,7 @@ class DoubleStreamBlock(nn.Module):
             attn = attention(torch.cat((img_q, txt_q), dim=2),
                              torch.cat((img_k, txt_k), dim=2),
                              torch.cat((img_v, txt_v), dim=2),
-                             pe=pe, mask=attn_mask)
+                             pe=pe, mask=attn_mask, transformer_options=transformer_options)
 
             img_attn, txt_attn = attn[:, : img.shape[1]], attn[:, img.shape[1]:]
         else:
@@ -190,7 +190,7 @@ class DoubleStreamBlock(nn.Module):
             attn = attention(torch.cat((txt_q, img_q), dim=2),
                              torch.cat((txt_k, img_k), dim=2),
                              torch.cat((txt_v, img_v), dim=2),
-                             pe=pe, mask=attn_mask)
+                             pe=pe, mask=attn_mask, transformer_options=transformer_options)
 
             txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
 
@@ -244,7 +244,7 @@ class SingleStreamBlock(nn.Module):
         self.mlp_act = nn.GELU(approximate="tanh")
         self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
 
-    def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims=None) -> Tensor:
+    def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims=None, transformer_options={}) -> Tensor:
         mod, _ = self.modulation(vec)
         qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
 
@@ -252,7 +252,7 @@ class SingleStreamBlock(nn.Module):
         q, k = self.norm(q, k, v)
 
         # compute attention
-        attn = attention(q, k, v, pe=pe, mask=attn_mask)
+        attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
         # compute activation in mlp stream, cat again and run second linear layer
         output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
         x += apply_mod(output, mod.gate, None, modulation_dims)

comfy/ldm/flux/math.py

@@ -6,7 +6,7 @@ from comfy.ldm.modules.attention import optimized_attention
 import comfy.model_management
 
 
-def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None) -> Tensor:
+def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor:
     q_shape = q.shape
     k_shape = k.shape
 
@@ -17,7 +17,7 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None) -> Tensor:
         k = (pe[..., 0] * k[..., 0] + pe[..., 1] * k[..., 1]).reshape(*k_shape).type_as(v)
 
     heads = q.shape[1]
-    x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask)
+    x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options)
     return x
 
 

comfy/ldm/flux/model.py

@@ -6,6 +6,7 @@ import torch
 from torch import Tensor, nn
 from einops import rearrange, repeat
 import comfy.ldm.common_dit
+import comfy.patcher_extension
 
 from .layers import (
     DoubleStreamBlock,
@@ -127,6 +128,7 @@ class Flux(nn.Module):
 
         blocks_replace = patches_replace.get("dit", {})
         for i, block in enumerate(self.double_blocks):
+            transformer_options["block"] = ("double_block", i, 2)
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -134,14 +136,16 @@ class Flux(nn.Module):
                                                    txt=args["txt"],
                                                    vec=args["vec"],
                                                    pe=args["pe"],
-                                                   attn_mask=args.get("attn_mask"))
+                                                   attn_mask=args.get("attn_mask"),
+                                                   transformer_options=args.get("transformer_options"))
                     return out
 
                 out = blocks_replace[("double_block", i)]({"img": img,
                                                            "txt": txt,
                                                            "vec": vec,
                                                            "pe": pe,
-                                                           "attn_mask": attn_mask},
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
                                                           {"original_block": block_wrap})
                 txt = out["txt"]
                 img = out["img"]
@@ -150,14 +154,15 @@ class Flux(nn.Module):
                                  txt=txt,
                                  vec=vec,
                                  pe=pe,
-                                 attn_mask=attn_mask)
+                                 attn_mask=attn_mask,
+                                 transformer_options=transformer_options)
 
             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[:, :add.shape[1]] += add
 
         if img.dtype == torch.float16:
             img = torch.nan_to_num(img, nan=0.0, posinf=65504, neginf=-65504)
@@ -165,30 +170,33 @@ class Flux(nn.Module):
         img = torch.cat((txt, img), 1)
 
         for i, block in enumerate(self.single_blocks):
+            transformer_options["block"] = ("single_block", i, 1)
             if ("single_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
                     out["img"] = block(args["img"],
                                        vec=args["vec"],
                                        pe=args["pe"],
-                                       attn_mask=args.get("attn_mask"))
+                                       attn_mask=args.get("attn_mask"),
+                                       transformer_options=args.get("transformer_options"))
                     return out
 
                 out = blocks_replace[("single_block", i)]({"img": img,
                                                            "vec": vec,
                                                            "pe": pe,
-                                                           "attn_mask": attn_mask},
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
                                                           {"original_block": block_wrap})
                 img = out["img"]
             else:
-                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, transformer_options=transformer_options)
 
             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[:, txt.shape[1] : txt.shape[1] + add.shape[1], ...] += add
 
         img = img[:, txt.shape[1] :, ...]
 
@@ -214,6 +222,13 @@ class Flux(nn.Module):
         return img, repeat(img_ids, "h w c -> b (h w) c", b=bs)
 
     def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, y, guidance, ref_latents, control, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs):
         bs, c, h_orig, w_orig = x.shape
         patch_size = self.patch_size
 
@@ -224,19 +239,27 @@ class Flux(nn.Module):
         if ref_latents is not None:
             h = 0
             w = 0
+            index = 0
+            index_ref_method = kwargs.get("ref_latents_method", "offset") == "index"
             for ref in ref_latents:
-                h_offset = 0
-                w_offset = 0
-                if ref.shape[-2] + h > ref.shape[-1] + w:
-                    w_offset = w
+                if index_ref_method:
+                    index += 1
+                    h_offset = 0
+                    w_offset = 0
                 else:
-                    h_offset = h
+                    index = 1
+                    h_offset = 0
+                    w_offset = 0
+                    if ref.shape[-2] + h > ref.shape[-1] + w:
+                        w_offset = w
+                    else:
+                        h_offset = h
+                    h = max(h, ref.shape[-2] + h_offset)
+                    w = max(w, ref.shape[-1] + w_offset)
 
-                kontext, kontext_ids = self.process_img(ref, index=1, h_offset=h_offset, w_offset=w_offset)
+                kontext, kontext_ids = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
                 img = torch.cat([img, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
-                h = max(h, ref.shape[-2] + h_offset)
-                w = max(w, ref.shape[-1] + w_offset)
 
         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, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))

comfy/ldm/genmo/joint_model/asymm_models_joint.py

@@ -109,6 +109,7 @@ class AsymmetricAttention(nn.Module):
         scale_x: torch.Tensor,  # (B, dim_x), modulation for pre-RMSNorm.
         scale_y: torch.Tensor,  # (B, dim_y), modulation for pre-RMSNorm.
         crop_y,
+        transformer_options={},
         **rope_rotation,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         rope_cos = rope_rotation.get("rope_cos")
@@ -143,7 +144,7 @@ class AsymmetricAttention(nn.Module):
 
         xy = optimized_attention(q,
                                  k,
-                                 v, self.num_heads, skip_reshape=True)
+                                 v, self.num_heads, skip_reshape=True, transformer_options=transformer_options)
 
         x, y = torch.tensor_split(xy, (q_x.shape[1],), dim=1)
         x = self.proj_x(x)
@@ -224,6 +225,7 @@ class AsymmetricJointBlock(nn.Module):
         x: torch.Tensor,
         c: torch.Tensor,
         y: torch.Tensor,
+        transformer_options={},
         **attn_kwargs,
     ):
         """Forward pass of a block.
@@ -256,6 +258,7 @@ class AsymmetricJointBlock(nn.Module):
             y,
             scale_x=scale_msa_x,
             scale_y=scale_msa_y,
+            transformer_options=transformer_options,
             **attn_kwargs,
         )
 
@@ -524,10 +527,11 @@ class AsymmDiTJoint(nn.Module):
                                                     args["txt"],
                                                     rope_cos=args["rope_cos"],
                                                     rope_sin=args["rope_sin"],
-                                                    crop_y=args["num_tokens"]
+                                                    crop_y=args["num_tokens"],
+                                                    transformer_options=args["transformer_options"]
                                                     )
                     return out
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": y_feat, "vec": c, "rope_cos": rope_cos, "rope_sin": rope_sin, "num_tokens": num_tokens}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": y_feat, "vec": c, "rope_cos": rope_cos, "rope_sin": rope_sin, "num_tokens": num_tokens, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 y_feat = out["txt"]
                 x = out["img"]
             else:
@@ -538,6 +542,7 @@ class AsymmDiTJoint(nn.Module):
                     rope_cos=rope_cos,
                     rope_sin=rope_sin,
                     crop_y=num_tokens,
+                    transformer_options=transformer_options,
                 )  # (B, M, D), (B, L, D)
         del y_feat  # Final layers don't use dense text features.
 

comfy/ldm/hidream/model.py

@@ -13,6 +13,7 @@ from comfy.ldm.flux.layers import LastLayer
 
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.model_management
+import comfy.patcher_extension
 import comfy.ldm.common_dit
 
 
@@ -71,8 +72,8 @@ class TimestepEmbed(nn.Module):
         return t_emb
 
 
-def attention(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor):
-    return optimized_attention(query.view(query.shape[0], -1, query.shape[-1] * query.shape[-2]), key.view(key.shape[0], -1, key.shape[-1] * key.shape[-2]), value.view(value.shape[0], -1, value.shape[-1] * value.shape[-2]), query.shape[2])
+def attention(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, transformer_options={}):
+    return optimized_attention(query.view(query.shape[0], -1, query.shape[-1] * query.shape[-2]), key.view(key.shape[0], -1, key.shape[-1] * key.shape[-2]), value.view(value.shape[0], -1, value.shape[-1] * value.shape[-2]), query.shape[2], transformer_options=transformer_options)
 
 
 class HiDreamAttnProcessor_flashattn:
@@ -85,6 +86,7 @@ class HiDreamAttnProcessor_flashattn:
         image_tokens_masks: Optional[torch.FloatTensor] = None,
         text_tokens: Optional[torch.FloatTensor] = None,
         rope: torch.FloatTensor = None,
+        transformer_options={},
         *args,
         **kwargs,
     ) -> torch.FloatTensor:
@@ -132,7 +134,7 @@ class HiDreamAttnProcessor_flashattn:
             query = torch.cat([query_1, query_2], dim=-1)
             key = torch.cat([key_1, key_2], dim=-1)
 
-        hidden_states = attention(query, key, value)
+        hidden_states = attention(query, key, value, transformer_options=transformer_options)
 
         if not attn.single:
             hidden_states_i, hidden_states_t = torch.split(hidden_states, [num_image_tokens, num_text_tokens], dim=1)
@@ -198,6 +200,7 @@ class HiDreamAttention(nn.Module):
         image_tokens_masks: torch.FloatTensor = None,
         norm_text_tokens: torch.FloatTensor = None,
         rope: torch.FloatTensor = None,
+        transformer_options={},
     ) -> torch.Tensor:
         return self.processor(
             self,
@@ -205,6 +208,7 @@ class HiDreamAttention(nn.Module):
             image_tokens_masks = image_tokens_masks,
             text_tokens = norm_text_tokens,
             rope = rope,
+            transformer_options=transformer_options,
         )
 
 
@@ -405,7 +409,7 @@ class HiDreamImageSingleTransformerBlock(nn.Module):
         text_tokens: Optional[torch.FloatTensor] = None,
         adaln_input: Optional[torch.FloatTensor] = None,
         rope: torch.FloatTensor = None,
-
+        transformer_options={},
     ) -> torch.FloatTensor:
         wtype = image_tokens.dtype
         shift_msa_i, scale_msa_i, gate_msa_i, shift_mlp_i, scale_mlp_i, gate_mlp_i = \
@@ -418,6 +422,7 @@ class HiDreamImageSingleTransformerBlock(nn.Module):
             norm_image_tokens,
             image_tokens_masks,
             rope = rope,
+            transformer_options=transformer_options,
         )
         image_tokens = gate_msa_i * attn_output_i + image_tokens
 
@@ -482,6 +487,7 @@ class HiDreamImageTransformerBlock(nn.Module):
         text_tokens: Optional[torch.FloatTensor] = None,
         adaln_input: Optional[torch.FloatTensor] = None,
         rope: torch.FloatTensor = None,
+        transformer_options={},
     ) -> torch.FloatTensor:
         wtype = image_tokens.dtype
         shift_msa_i, scale_msa_i, gate_msa_i, shift_mlp_i, scale_mlp_i, gate_mlp_i, \
@@ -499,6 +505,7 @@ class HiDreamImageTransformerBlock(nn.Module):
             image_tokens_masks,
             norm_text_tokens,
             rope = rope,
+            transformer_options=transformer_options,
         )
 
         image_tokens = gate_msa_i * attn_output_i + image_tokens
@@ -549,6 +556,7 @@ class HiDreamImageBlock(nn.Module):
         text_tokens: Optional[torch.FloatTensor] = None,
         adaln_input: torch.FloatTensor = None,
         rope: torch.FloatTensor = None,
+        transformer_options={},
     ) -> torch.FloatTensor:
         return self.block(
             image_tokens,
@@ -556,6 +564,7 @@ class HiDreamImageBlock(nn.Module):
             text_tokens,
             adaln_input,
             rope,
+            transformer_options=transformer_options,
         )
 
 
@@ -692,7 +701,23 @@ class HiDreamImageTransformer2DModel(nn.Module):
             raise NotImplementedError
         return x, x_masks, img_sizes
 
-    def forward(
+    def forward(self,
+        x: torch.Tensor,
+        t: torch.Tensor,
+        y: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        encoder_hidden_states_llama3=None,
+        image_cond=None,
+        control = None,
+        transformer_options = {},
+    ):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, t, y, context, encoder_hidden_states_llama3, image_cond, control, transformer_options)
+
+    def _forward(
         self,
         x: torch.Tensor,
         t: torch.Tensor,
@@ -769,6 +794,7 @@ class HiDreamImageTransformer2DModel(nn.Module):
                 text_tokens = cur_encoder_hidden_states,
                 adaln_input = adaln_input,
                 rope = rope,
+                transformer_options=transformer_options,
             )
             initial_encoder_hidden_states = initial_encoder_hidden_states[:, :initial_encoder_hidden_states_seq_len]
             block_id += 1
@@ -792,6 +818,7 @@ class HiDreamImageTransformer2DModel(nn.Module):
                 text_tokens=None,
                 adaln_input=adaln_input,
                 rope=rope,
+                transformer_options=transformer_options,
             )
             hidden_states = hidden_states[:, :hidden_states_seq_len]
             block_id += 1

comfy/ldm/hunyuan3d/model.py

@@ -7,6 +7,7 @@ from comfy.ldm.flux.layers import (
     SingleStreamBlock,
     timestep_embedding,
 )
+import comfy.patcher_extension
 
 
 class Hunyuan3Dv2(nn.Module):
@@ -67,6 +68,13 @@ class Hunyuan3Dv2(nn.Module):
         self.final_layer = LastLayer(hidden_size, 1, in_channels, dtype=dtype, device=device, operations=operations)
 
     def forward(self, x, timestep, context, guidance=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, guidance, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, guidance=None, transformer_options={}, **kwargs):
         x = x.movedim(-1, -2)
         timestep = 1.0 - timestep
         txt = context
@@ -91,14 +99,16 @@ class Hunyuan3Dv2(nn.Module):
                                                    txt=args["txt"],
                                                    vec=args["vec"],
                                                    pe=args["pe"],
-                                                   attn_mask=args.get("attn_mask"))
+                                                   attn_mask=args.get("attn_mask"),
+                                                   transformer_options=args["transformer_options"])
                     return out
 
                 out = blocks_replace[("double_block", i)]({"img": img,
                                                            "txt": txt,
                                                            "vec": vec,
                                                            "pe": pe,
-                                                           "attn_mask": attn_mask},
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
                                                           {"original_block": block_wrap})
                 txt = out["txt"]
                 img = out["img"]
@@ -107,7 +117,8 @@ class Hunyuan3Dv2(nn.Module):
                                  txt=txt,
                                  vec=vec,
                                  pe=pe,
-                                 attn_mask=attn_mask)
+                                 attn_mask=attn_mask,
+                                 transformer_options=transformer_options)
 
         img = torch.cat((txt, img), 1)
 
@@ -118,17 +129,19 @@ class Hunyuan3Dv2(nn.Module):
                     out["img"] = block(args["img"],
                                        vec=args["vec"],
                                        pe=args["pe"],
-                                       attn_mask=args.get("attn_mask"))
+                                       attn_mask=args.get("attn_mask"),
+                                       transformer_options=args["transformer_options"])
                     return out
 
                 out = blocks_replace[("single_block", i)]({"img": img,
                                                            "vec": vec,
                                                            "pe": pe,
-                                                           "attn_mask": attn_mask},
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
                                                           {"original_block": block_wrap})
                 img = out["img"]
             else:
-                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, transformer_options=transformer_options)
 
         img = img[:, txt.shape[1]:, ...]
         img = self.final_layer(img, vec)

comfy/ldm/hunyuan3d/vae.py

@@ -178,7 +178,7 @@ class FourierEmbedder(nn.Module):
 
 class CrossAttentionProcessor:
     def __call__(self, attn, q, k, v):
-        out = F.scaled_dot_product_attention(q, k, v)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v)
         return out
 
 

comfy/ldm/hunyuan_video/model.py

@@ -1,6 +1,7 @@
 #Based on Flux code because of weird hunyuan video code license.
 
 import torch
+import comfy.patcher_extension
 import comfy.ldm.flux.layers
 import comfy.ldm.modules.diffusionmodules.mmdit
 from comfy.ldm.modules.attention import optimized_attention
@@ -77,13 +78,13 @@ class TokenRefinerBlock(nn.Module):
             operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
         )
 
-    def forward(self, x, c, mask):
+    def forward(self, x, c, mask, transformer_options={}):
         mod1, mod2 = self.adaLN_modulation(c).chunk(2, dim=1)
 
         norm_x = self.norm1(x)
         qkv = self.self_attn.qkv(norm_x)
         q, k, v = qkv.reshape(qkv.shape[0], qkv.shape[1], 3, self.heads, -1).permute(2, 0, 3, 1, 4)
-        attn = optimized_attention(q, k, v, self.heads, mask=mask, skip_reshape=True)
+        attn = optimized_attention(q, k, v, self.heads, mask=mask, skip_reshape=True, transformer_options=transformer_options)
 
         x = x + self.self_attn.proj(attn) * mod1.unsqueeze(1)
         x = x + self.mlp(self.norm2(x)) * mod2.unsqueeze(1)
@@ -114,14 +115,14 @@ class IndividualTokenRefiner(nn.Module):
             ]
         )
 
-    def forward(self, x, c, mask):
+    def forward(self, x, c, mask, transformer_options={}):
         m = None
         if mask is not None:
             m = mask.view(mask.shape[0], 1, 1, mask.shape[1]).repeat(1, 1, mask.shape[1], 1)
             m = m + m.transpose(2, 3)
 
         for block in self.blocks:
-            x = block(x, c, m)
+            x = block(x, c, m, transformer_options=transformer_options)
         return x
 
 
@@ -149,6 +150,7 @@ class TokenRefiner(nn.Module):
         x,
         timesteps,
         mask,
+        transformer_options={},
     ):
         t = self.t_embedder(timestep_embedding(timesteps, 256, time_factor=1.0).to(x.dtype))
         # m = mask.float().unsqueeze(-1)
@@ -157,7 +159,7 @@ class TokenRefiner(nn.Module):
 
         c = t + self.c_embedder(c.to(x.dtype))
         x = self.input_embedder(x)
-        x = self.individual_token_refiner(x, c, mask)
+        x = self.individual_token_refiner(x, c, mask, transformer_options=transformer_options)
         return x
 
 class HunyuanVideo(nn.Module):
@@ -266,7 +268,7 @@ class HunyuanVideo(nn.Module):
         if txt_mask is not None and not torch.is_floating_point(txt_mask):
             txt_mask = (txt_mask - 1).to(img.dtype) * torch.finfo(img.dtype).max
 
-        txt = self.txt_in(txt, timesteps, txt_mask)
+        txt = self.txt_in(txt, timesteps, txt_mask, transformer_options=transformer_options)
 
         ids = torch.cat((img_ids, txt_ids), dim=1)
         pe = self.pe_embedder(ids)
@@ -284,14 +286,14 @@ class HunyuanVideo(nn.Module):
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"], out["txt"] = block(img=args["img"], txt=args["txt"], vec=args["vec"], pe=args["pe"], attn_mask=args["attention_mask"], modulation_dims_img=args["modulation_dims_img"], modulation_dims_txt=args["modulation_dims_txt"])
+                    out["img"], out["txt"] = block(img=args["img"], txt=args["txt"], vec=args["vec"], pe=args["pe"], attn_mask=args["attention_mask"], modulation_dims_img=args["modulation_dims_img"], modulation_dims_txt=args["modulation_dims_txt"], transformer_options=args["transformer_options"])
                     return out
 
-                out = blocks_replace[("double_block", i)]({"img": img, "txt": txt, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims_img': modulation_dims, 'modulation_dims_txt': modulation_dims_txt}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": img, "txt": txt, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims_img': modulation_dims, 'modulation_dims_txt': modulation_dims_txt, 'transformer_options': transformer_options}, {"original_block": block_wrap})
                 txt = out["txt"]
                 img = out["img"]
             else:
-                img, txt = block(img=img, txt=txt, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims_img=modulation_dims, modulation_dims_txt=modulation_dims_txt)
+                img, txt = block(img=img, txt=txt, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims_img=modulation_dims, modulation_dims_txt=modulation_dims_txt, transformer_options=transformer_options)
 
             if control is not None: # Controlnet
                 control_i = control.get("input")
@@ -306,13 +308,13 @@ class HunyuanVideo(nn.Module):
             if ("single_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"] = block(args["img"], vec=args["vec"], pe=args["pe"], attn_mask=args["attention_mask"], modulation_dims=args["modulation_dims"])
+                    out["img"] = block(args["img"], vec=args["vec"], pe=args["pe"], attn_mask=args["attention_mask"], modulation_dims=args["modulation_dims"], transformer_options=args["transformer_options"])
                     return out
 
-                out = blocks_replace[("single_block", i)]({"img": img, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims': modulation_dims}, {"original_block": block_wrap})
+                out = blocks_replace[("single_block", i)]({"img": img, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims': modulation_dims, 'transformer_options': transformer_options}, {"original_block": block_wrap})
                 img = out["img"]
             else:
-                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims=modulation_dims)
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims=modulation_dims, transformer_options=transformer_options)
 
             if control is not None: # Controlnet
                 control_o = control.get("output")
@@ -348,6 +350,13 @@ class HunyuanVideo(nn.Module):
         return repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
 
     def forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, control=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, y, guidance, attention_mask, guiding_frame_index, ref_latent, control, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, control=None, transformer_options={}, **kwargs):
         bs, c, t, h, w = x.shape
         img_ids = self.img_ids(x)
         txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)

comfy/ldm/lightricks/model.py

@@ -1,5 +1,6 @@
 import torch
 from torch import nn
+import comfy.patcher_extension
 import comfy.ldm.modules.attention
 import comfy.ldm.common_dit
 from einops import rearrange
@@ -270,7 +271,7 @@ class CrossAttention(nn.Module):
 
         self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout))
 
-    def forward(self, x, context=None, mask=None, pe=None):
+    def forward(self, x, context=None, mask=None, pe=None, transformer_options={}):
         q = self.to_q(x)
         context = x if context is None else context
         k = self.to_k(context)
@@ -284,9 +285,9 @@ class CrossAttention(nn.Module):
             k = apply_rotary_emb(k, pe)
 
         if mask is None:
-            out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision)
+            out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision, transformer_options=transformer_options)
         else:
-            out = comfy.ldm.modules.attention.optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision)
+            out = comfy.ldm.modules.attention.optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision, transformer_options=transformer_options)
         return self.to_out(out)
 
 
@@ -302,12 +303,12 @@ class BasicTransformerBlock(nn.Module):
 
         self.scale_shift_table = nn.Parameter(torch.empty(6, dim, device=device, dtype=dtype))
 
-    def forward(self, x, context=None, attention_mask=None, timestep=None, pe=None):
+    def forward(self, x, context=None, attention_mask=None, timestep=None, pe=None, transformer_options={}):
         shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + timestep.reshape(x.shape[0], timestep.shape[1], self.scale_shift_table.shape[0], -1)).unbind(dim=2)
 
-        x += self.attn1(comfy.ldm.common_dit.rms_norm(x) * (1 + scale_msa) + shift_msa, pe=pe) * gate_msa
+        x += self.attn1(comfy.ldm.common_dit.rms_norm(x) * (1 + scale_msa) + shift_msa, pe=pe, transformer_options=transformer_options) * gate_msa
 
-        x += self.attn2(x, context=context, mask=attention_mask)
+        x += self.attn2(x, context=context, mask=attention_mask, transformer_options=transformer_options)
 
         y = comfy.ldm.common_dit.rms_norm(x) * (1 + scale_mlp) + shift_mlp
         x += self.ff(y) * gate_mlp
@@ -420,6 +421,13 @@ class LTXVModel(torch.nn.Module):
         self.patchifier = SymmetricPatchifier(1)
 
     def forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, **kwargs)
+
+    def _forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
         patches_replace = transformer_options.get("patches_replace", {})
 
         orig_shape = list(x.shape)
@@ -471,10 +479,10 @@ class LTXVModel(torch.nn.Module):
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"] = block(args["img"], context=args["txt"], attention_mask=args["attention_mask"], timestep=args["vec"], pe=args["pe"])
+                    out["img"] = block(args["img"], context=args["txt"], attention_mask=args["attention_mask"], timestep=args["vec"], pe=args["pe"], transformer_options=args["transformer_options"])
                     return out
 
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "attention_mask": attention_mask, "vec": timestep, "pe": pe}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "attention_mask": attention_mask, "vec": timestep, "pe": pe, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 x = out["img"]
             else:
                 x = block(
@@ -482,7 +490,8 @@ class LTXVModel(torch.nn.Module):
                     context=context,
                     attention_mask=attention_mask,
                     timestep=timestep,
-                    pe=pe
+                    pe=pe,
+                    transformer_options=transformer_options,
                 )
 
         # 3. Output

comfy/ldm/lumina/model.py

@@ -11,6 +11,7 @@ import comfy.ldm.common_dit
 from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder
 from comfy.ldm.modules.attention import optimized_attention_masked
 from comfy.ldm.flux.layers import EmbedND
+import comfy.patcher_extension
 
 
 def modulate(x, scale):
@@ -103,6 +104,7 @@ class JointAttention(nn.Module):
         x: torch.Tensor,
         x_mask: torch.Tensor,
         freqs_cis: torch.Tensor,
+        transformer_options={},
     ) -> torch.Tensor:
         """
 
@@ -139,7 +141,7 @@ class JointAttention(nn.Module):
         if n_rep >= 1:
             xk = xk.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
             xv = xv.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
-        output = optimized_attention_masked(xq.movedim(1, 2), xk.movedim(1, 2), xv.movedim(1, 2), self.n_local_heads, x_mask, skip_reshape=True)
+        output = optimized_attention_masked(xq.movedim(1, 2), xk.movedim(1, 2), xv.movedim(1, 2), self.n_local_heads, x_mask, skip_reshape=True, transformer_options=transformer_options)
 
         return self.out(output)
 
@@ -267,6 +269,7 @@ class JointTransformerBlock(nn.Module):
         x_mask: torch.Tensor,
         freqs_cis: torch.Tensor,
         adaln_input: Optional[torch.Tensor]=None,
+        transformer_options={},
     ):
         """
         Perform a forward pass through the TransformerBlock.
@@ -289,6 +292,7 @@ class JointTransformerBlock(nn.Module):
                     modulate(self.attention_norm1(x), scale_msa),
                     x_mask,
                     freqs_cis,
+                    transformer_options=transformer_options,
                 )
             )
             x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(
@@ -303,6 +307,7 @@ class JointTransformerBlock(nn.Module):
                     self.attention_norm1(x),
                     x_mask,
                     freqs_cis,
+                    transformer_options=transformer_options,
                 )
             )
             x = x + self.ffn_norm2(
@@ -493,7 +498,7 @@ class NextDiT(nn.Module):
         return imgs
 
     def patchify_and_embed(
-        self, x: List[torch.Tensor] | torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens
+        self, x: List[torch.Tensor] | torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens, transformer_options={}
     ) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], List[int], torch.Tensor]:
         bsz = len(x)
         pH = pW = self.patch_size
@@ -553,7 +558,7 @@ class NextDiT(nn.Module):
 
         # refine context
         for layer in self.context_refiner:
-            cap_feats = layer(cap_feats, cap_mask, cap_freqs_cis)
+            cap_feats = layer(cap_feats, cap_mask, cap_freqs_cis, transformer_options=transformer_options)
 
         # refine image
         flat_x = []
@@ -572,7 +577,7 @@ class NextDiT(nn.Module):
         padded_img_embed = self.x_embedder(padded_img_embed)
         padded_img_mask = padded_img_mask.unsqueeze(1)
         for layer in self.noise_refiner:
-            padded_img_embed = layer(padded_img_embed, padded_img_mask, img_freqs_cis, t)
+            padded_img_embed = layer(padded_img_embed, padded_img_mask, img_freqs_cis, t, transformer_options=transformer_options)
 
         if cap_mask is not None:
             mask = torch.zeros(bsz, max_seq_len, dtype=dtype, device=device)
@@ -590,8 +595,15 @@ class NextDiT(nn.Module):
 
         return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis
 
-    # def forward(self, x, t, cap_feats, cap_mask):
     def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, kwargs.get("transformer_options", {}))
+        ).execute(x, timesteps, context, num_tokens, attention_mask, **kwargs)
+
+    # def forward(self, x, t, cap_feats, cap_mask):
+    def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
         t = 1.0 - timesteps
         cap_feats = context
         cap_mask = attention_mask
@@ -608,12 +620,13 @@ class NextDiT(nn.Module):
 
         cap_feats = self.cap_embedder(cap_feats)  # (N, L, D)  # todo check if able to batchify w.o. redundant compute
 
+        transformer_options = kwargs.get("transformer_options", {})
         x_is_tensor = isinstance(x, torch.Tensor)
-        x, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens)
+        x, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens, transformer_options=transformer_options)
         freqs_cis = freqs_cis.to(x.device)
 
         for layer in self.layers:
-            x = layer(x, mask, freqs_cis, adaln_input)
+            x = layer(x, mask, freqs_cis, adaln_input, transformer_options=transformer_options)
 
         x = self.final_layer(x, adaln_input)
         x = self.unpatchify(x, img_size, cap_size, return_tensor=x_is_tensor)[:,:,:h,:w]

comfy/ldm/modules/attention.py

@@ -5,8 +5,9 @@ import torch
 import torch.nn.functional as F
 from torch import nn, einsum
 from einops import rearrange, repeat
-from typing import Optional
+from typing import Optional, Any, Callable, Union
 import logging
+import functools
 
 from .diffusionmodules.util import AlphaBlender, timestep_embedding
 from .sub_quadratic_attention import efficient_dot_product_attention
@@ -17,23 +18,45 @@ if model_management.xformers_enabled():
     import xformers
     import xformers.ops
 
-if model_management.sage_attention_enabled():
-    try:
-        from sageattention import sageattn
-    except ModuleNotFoundError as e:
+SAGE_ATTENTION_IS_AVAILABLE = False
+try:
+    from sageattention import sageattn
+    SAGE_ATTENTION_IS_AVAILABLE = True
+except ModuleNotFoundError as e:
+    if model_management.sage_attention_enabled():
         if e.name == "sageattention":
             logging.error(f"\n\nTo use the `--use-sage-attention` feature, the `sageattention` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install sageattention")
         else:
             raise e
         exit(-1)
 
-if model_management.flash_attention_enabled():
-    try:
-        from flash_attn import flash_attn_func
-    except ModuleNotFoundError:
+FLASH_ATTENTION_IS_AVAILABLE = False
+try:
+    from flash_attn import flash_attn_func
+    FLASH_ATTENTION_IS_AVAILABLE = True
+except ModuleNotFoundError:
+    if model_management.flash_attention_enabled():
         logging.error(f"\n\nTo use the `--use-flash-attention` feature, the `flash-attn` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install flash-attn")
         exit(-1)
 
+REGISTERED_ATTENTION_FUNCTIONS = {}
+def register_attention_function(name: str, func: Callable):
+    # avoid replacing existing functions
+    if name not in REGISTERED_ATTENTION_FUNCTIONS:
+        REGISTERED_ATTENTION_FUNCTIONS[name] = func
+    else:
+        logging.warning(f"Attention function {name} already registered, skipping registration.")
+
+def get_attention_function(name: str, default: Any=...) -> Union[Callable, None]:
+    if name == "optimized":
+        return optimized_attention
+    elif name not in REGISTERED_ATTENTION_FUNCTIONS:
+        if default is ...:
+            raise KeyError(f"Attention function {name} not found.")
+        else:
+            return default
+    return REGISTERED_ATTENTION_FUNCTIONS[name]
+
 from comfy.cli_args import args
 import comfy.ops
 ops = comfy.ops.disable_weight_init
@@ -91,7 +114,27 @@ class FeedForward(nn.Module):
 def Normalize(in_channels, dtype=None, device=None):
     return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
 
-def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+
+def wrap_attn(func):
+    @functools.wraps(func)
+    def wrapper(*args, **kwargs):
+        remove_attn_wrapper_key = False
+        try:
+            if "_inside_attn_wrapper" not in kwargs:
+                transformer_options = kwargs.get("transformer_options", None)
+                remove_attn_wrapper_key = True
+                kwargs["_inside_attn_wrapper"] = True
+                if transformer_options is not None:
+                    if "optimized_attention_override" in transformer_options:
+                        return transformer_options["optimized_attention_override"](func, *args, **kwargs)
+            return func(*args, **kwargs)
+        finally:
+            if remove_attn_wrapper_key:
+                del kwargs["_inside_attn_wrapper"]
+    return wrapper
+
+@wrap_attn
+def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     attn_precision = get_attn_precision(attn_precision, q.dtype)
 
     if skip_reshape:
@@ -159,8 +202,8 @@ def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
         )
     return out
 
-
-def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+@wrap_attn
+def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     attn_precision = get_attn_precision(attn_precision, query.dtype)
 
     if skip_reshape:
@@ -230,7 +273,8 @@ def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None,
         hidden_states = hidden_states.unflatten(0, (-1, heads)).transpose(1,2).flatten(start_dim=2)
     return hidden_states
 
-def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+@wrap_attn
+def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     attn_precision = get_attn_precision(attn_precision, q.dtype)
 
     if skip_reshape:
@@ -359,7 +403,8 @@ try:
 except:
     pass
 
-def attention_xformers(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+@wrap_attn
+def attention_xformers(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     b = q.shape[0]
     dim_head = q.shape[-1]
     # check to make sure xformers isn't broken
@@ -374,7 +419,7 @@ 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, skip_reshape=skip_reshape)
+        return attention_pytorch(q, k, v, heads, mask, skip_reshape=skip_reshape, **kwargs)
 
     if skip_reshape:
         # b h k d -> b k h d
@@ -427,8 +472,8 @@ else:
     #TODO: other GPUs ?
     SDP_BATCH_LIMIT = 2**31
 
-
-def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+@wrap_attn
+def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     if skip_reshape:
         b, _, _, dim_head = q.shape
     else:
@@ -448,7 +493,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
             mask = mask.unsqueeze(1)
 
     if SDP_BATCH_LIMIT >= b:
-        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
         if not skip_output_reshape:
             out = (
                 out.transpose(1, 2).reshape(b, -1, heads * dim_head)
@@ -461,7 +506,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
                 if mask.shape[0] > 1:
                     m = mask[i : i + SDP_BATCH_LIMIT]
 
-            out[i : i + SDP_BATCH_LIMIT] = torch.nn.functional.scaled_dot_product_attention(
+            out[i : i + SDP_BATCH_LIMIT] = comfy.ops.scaled_dot_product_attention(
                 q[i : i + SDP_BATCH_LIMIT],
                 k[i : i + SDP_BATCH_LIMIT],
                 v[i : i + SDP_BATCH_LIMIT],
@@ -470,8 +515,8 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
             ).transpose(1, 2).reshape(-1, q.shape[2], heads * dim_head)
     return out
 
-
-def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+@wrap_attn
+def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     if skip_reshape:
         b, _, _, dim_head = q.shape
         tensor_layout = "HND"
@@ -501,7 +546,7 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
                 lambda t: t.transpose(1, 2),
                 (q, k, v),
             )
-        return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=True, skip_output_reshape=skip_output_reshape)
+        return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=True, skip_output_reshape=skip_output_reshape, **kwargs)
 
     if tensor_layout == "HND":
         if not skip_output_reshape:
@@ -534,8 +579,8 @@ except AttributeError as error:
                     dropout_p: float = 0.0, causal: bool = False) -> torch.Tensor:
         assert False, f"Could not define flash_attn_wrapper: {FLASH_ATTN_ERROR}"
 
-
-def attention_flash(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+@wrap_attn
+def attention_flash(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
     if skip_reshape:
         b, _, _, dim_head = q.shape
     else:
@@ -597,6 +642,19 @@ else:
 
 optimized_attention_masked = optimized_attention
 
+
+# register core-supported attention functions
+if SAGE_ATTENTION_IS_AVAILABLE:
+    register_attention_function("sage", attention_sage)
+if FLASH_ATTENTION_IS_AVAILABLE:
+    register_attention_function("flash", attention_flash)
+if model_management.xformers_enabled():
+    register_attention_function("xformers", attention_xformers)
+register_attention_function("pytorch", attention_pytorch)
+register_attention_function("sub_quad", attention_sub_quad)
+register_attention_function("split", attention_split)
+
+
 def optimized_attention_for_device(device, mask=False, small_input=False):
     if small_input:
         if model_management.pytorch_attention_enabled():
@@ -629,7 +687,7 @@ class CrossAttention(nn.Module):
 
         self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout))
 
-    def forward(self, x, context=None, value=None, mask=None):
+    def forward(self, x, context=None, value=None, mask=None, transformer_options={}):
         q = self.to_q(x)
         context = default(context, x)
         k = self.to_k(context)
@@ -640,9 +698,9 @@ class CrossAttention(nn.Module):
             v = self.to_v(context)
 
         if mask is None:
-            out = optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision)
+            out = optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision, transformer_options=transformer_options)
         else:
-            out = optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision)
+            out = optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision, transformer_options=transformer_options)
         return self.to_out(out)
 
 
@@ -746,7 +804,7 @@ class BasicTransformerBlock(nn.Module):
             n = attn1_replace_patch[block_attn1](n, context_attn1, value_attn1, extra_options)
             n = self.attn1.to_out(n)
         else:
-            n = self.attn1(n, context=context_attn1, value=value_attn1)
+            n = self.attn1(n, context=context_attn1, value=value_attn1, transformer_options=transformer_options)
 
         if "attn1_output_patch" in transformer_patches:
             patch = transformer_patches["attn1_output_patch"]
@@ -786,7 +844,7 @@ class BasicTransformerBlock(nn.Module):
                 n = attn2_replace_patch[block_attn2](n, context_attn2, value_attn2, extra_options)
                 n = self.attn2.to_out(n)
             else:
-                n = self.attn2(n, context=context_attn2, value=value_attn2)
+                n = self.attn2(n, context=context_attn2, value=value_attn2, transformer_options=transformer_options)
 
         if "attn2_output_patch" in transformer_patches:
             patch = transformer_patches["attn2_output_patch"]
@@ -1017,7 +1075,7 @@ class SpatialVideoTransformer(SpatialTransformer):
 
             B, S, C = x_mix.shape
             x_mix = rearrange(x_mix, "(b t) s c -> (b s) t c", t=timesteps)
-            x_mix = mix_block(x_mix, context=time_context) #TODO: transformer_options
+            x_mix = mix_block(x_mix, context=time_context, transformer_options=transformer_options)
             x_mix = rearrange(
                 x_mix, "(b s) t c -> (b t) s c", s=S, b=B // timesteps, c=C, t=timesteps
             )

comfy/ldm/modules/diffusionmodules/mmdit.py

@@ -109,7 +109,7 @@ class PatchEmbed(nn.Module):
 def modulate(x, shift, scale):
     if shift is None:
         shift = torch.zeros_like(scale)
-    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+    return torch.addcmul(shift.unsqueeze(1), x, 1+ scale.unsqueeze(1))
 
 
 #################################################################################
@@ -564,10 +564,7 @@ class DismantledBlock(nn.Module):
         assert not self.pre_only
         attn1 = self.attn.post_attention(attn)
         attn2 = self.attn2.post_attention(attn2)
-        out1 = gate_msa.unsqueeze(1) * attn1
-        out2 = gate_msa2.unsqueeze(1) * attn2
-        x = x + out1
-        x = x + out2
+        x = gate_cat(x, gate_msa, gate_msa2, attn1, attn2)
         x = x + gate_mlp.unsqueeze(1) * self.mlp(
             modulate(self.norm2(x), shift_mlp, scale_mlp)
         )
@@ -594,6 +591,11 @@ class DismantledBlock(nn.Module):
             )
             return self.post_attention(attn, *intermediates)
 
+def gate_cat(x, gate_msa, gate_msa2, attn1, attn2):
+    out1 = gate_msa.unsqueeze(1) * attn1
+    out2 = gate_msa2.unsqueeze(1) * attn2
+    x = torch.stack([x, out1, out2], dim=0).sum(dim=0)
+    return x
 
 def block_mixing(*args, use_checkpoint=True, **kwargs):
     if use_checkpoint:
@@ -604,7 +606,7 @@ def block_mixing(*args, use_checkpoint=True, **kwargs):
         return _block_mixing(*args, **kwargs)
 
 
-def _block_mixing(context, x, context_block, x_block, c):
+def _block_mixing(context, x, context_block, x_block, c, transformer_options={}):
     context_qkv, context_intermediates = context_block.pre_attention(context, c)
 
     if x_block.x_block_self_attn:
@@ -620,6 +622,7 @@ def _block_mixing(context, x, context_block, x_block, c):
     attn = optimized_attention(
         qkv[0], qkv[1], qkv[2],
         heads=x_block.attn.num_heads,
+        transformer_options=transformer_options,
     )
     context_attn, x_attn = (
         attn[:, : context_qkv[0].shape[1]],
@@ -635,6 +638,7 @@ def _block_mixing(context, x, context_block, x_block, c):
         attn2 = optimized_attention(
                 x_qkv2[0], x_qkv2[1], x_qkv2[2],
                 heads=x_block.attn2.num_heads,
+                transformer_options=transformer_options,
             )
         x = x_block.post_attention_x(x_attn, attn2, *x_intermediates)
     else:
@@ -956,10 +960,10 @@ class MMDiT(nn.Module):
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["txt"], out["img"] = self.joint_blocks[i](args["txt"], args["img"], c=args["vec"])
+                    out["txt"], out["img"] = self.joint_blocks[i](args["txt"], args["img"], c=args["vec"], transformer_options=args["transformer_options"])
                     return out
 
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": c_mod}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": c_mod, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 context = out["txt"]
                 x = out["img"]
             else:
@@ -968,6 +972,7 @@ class MMDiT(nn.Module):
                     x,
                     c=c_mod,
                     use_checkpoint=self.use_checkpoint,
+                    transformer_options=transformer_options,
                 )
             if control is not None:
                 control_o = control.get("output")

comfy/ldm/modules/diffusionmodules/model.py

@@ -285,7 +285,7 @@ def pytorch_attention(q, k, v):
     )
 
     try:
-        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
         out = out.transpose(2, 3).reshape(orig_shape)
     except model_management.OOM_EXCEPTION:
         logging.warning("scaled_dot_product_attention OOMed: switched to slice attention")

comfy/ldm/omnigen/omnigen2.py

@@ -120,7 +120,7 @@ class Attention(nn.Module):
             nn.Dropout(0.0)
         )
 
-    def forward(self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, image_rotary_emb: Optional[torch.Tensor] = None) -> torch.Tensor:
+    def forward(self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, image_rotary_emb: Optional[torch.Tensor] = None, transformer_options={}) -> torch.Tensor:
         batch_size, sequence_length, _ = hidden_states.shape
 
         query = self.to_q(hidden_states)
@@ -146,7 +146,7 @@ class Attention(nn.Module):
             key = key.repeat_interleave(self.heads // self.kv_heads, dim=1)
             value = value.repeat_interleave(self.heads // self.kv_heads, dim=1)
 
-        hidden_states = optimized_attention_masked(query, key, value, self.heads, attention_mask, skip_reshape=True)
+        hidden_states = optimized_attention_masked(query, key, value, self.heads, attention_mask, skip_reshape=True, transformer_options=transformer_options)
         hidden_states = self.to_out[0](hidden_states)
         return hidden_states
 
@@ -182,16 +182,16 @@ class OmniGen2TransformerBlock(nn.Module):
         self.norm2 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
         self.ffn_norm2 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
 
-    def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, image_rotary_emb: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor:
+    def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, image_rotary_emb: torch.Tensor, temb: Optional[torch.Tensor] = None, transformer_options={}) -> torch.Tensor:
         if self.modulation:
             norm_hidden_states, gate_msa, scale_mlp, gate_mlp = self.norm1(hidden_states, temb)
-            attn_output = self.attn(norm_hidden_states, norm_hidden_states, attention_mask, image_rotary_emb)
+            attn_output = self.attn(norm_hidden_states, norm_hidden_states, attention_mask, image_rotary_emb, transformer_options=transformer_options)
             hidden_states = hidden_states + gate_msa.unsqueeze(1).tanh() * self.norm2(attn_output)
             mlp_output = self.feed_forward(self.ffn_norm1(hidden_states) * (1 + scale_mlp.unsqueeze(1)))
             hidden_states = hidden_states + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(mlp_output)
         else:
             norm_hidden_states = self.norm1(hidden_states)
-            attn_output = self.attn(norm_hidden_states, norm_hidden_states, attention_mask, image_rotary_emb)
+            attn_output = self.attn(norm_hidden_states, norm_hidden_states, attention_mask, image_rotary_emb, transformer_options=transformer_options)
             hidden_states = hidden_states + self.norm2(attn_output)
             mlp_output = self.feed_forward(self.ffn_norm1(hidden_states))
             hidden_states = hidden_states + self.ffn_norm2(mlp_output)
@@ -390,7 +390,7 @@ class OmniGen2Transformer2DModel(nn.Module):
             ref_img_sizes, img_sizes,
         )
 
-    def img_patch_embed_and_refine(self, hidden_states, ref_image_hidden_states, padded_img_mask, padded_ref_img_mask, noise_rotary_emb, ref_img_rotary_emb, l_effective_ref_img_len, l_effective_img_len, temb):
+    def img_patch_embed_and_refine(self, hidden_states, ref_image_hidden_states, padded_img_mask, padded_ref_img_mask, noise_rotary_emb, ref_img_rotary_emb, l_effective_ref_img_len, l_effective_img_len, temb, transformer_options={}):
         batch_size = len(hidden_states)
 
         hidden_states = self.x_embedder(hidden_states)
@@ -405,17 +405,17 @@ class OmniGen2Transformer2DModel(nn.Module):
                     shift += ref_img_len
 
         for layer in self.noise_refiner:
-            hidden_states = layer(hidden_states, padded_img_mask, noise_rotary_emb, temb)
+            hidden_states = layer(hidden_states, padded_img_mask, noise_rotary_emb, temb, transformer_options=transformer_options)
 
         if ref_image_hidden_states is not None:
             for layer in self.ref_image_refiner:
-                ref_image_hidden_states = layer(ref_image_hidden_states, padded_ref_img_mask, ref_img_rotary_emb, temb)
+                ref_image_hidden_states = layer(ref_image_hidden_states, padded_ref_img_mask, ref_img_rotary_emb, temb, transformer_options=transformer_options)
 
             hidden_states = torch.cat([ref_image_hidden_states, hidden_states], dim=1)
 
         return hidden_states
 
-    def forward(self, x, timesteps, context, num_tokens, ref_latents=None, attention_mask=None, **kwargs):
+    def forward(self, x, timesteps, context, num_tokens, ref_latents=None, attention_mask=None, transformer_options={}, **kwargs):
         B, C, H, W = x.shape
         hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
         _, _, H_padded, W_padded = hidden_states.shape
@@ -444,7 +444,7 @@ class OmniGen2Transformer2DModel(nn.Module):
         )
 
         for layer in self.context_refiner:
-            text_hidden_states = layer(text_hidden_states, text_attention_mask, context_rotary_emb)
+            text_hidden_states = layer(text_hidden_states, text_attention_mask, context_rotary_emb, transformer_options=transformer_options)
 
         img_len = hidden_states.shape[1]
         combined_img_hidden_states = self.img_patch_embed_and_refine(
@@ -453,13 +453,14 @@ class OmniGen2Transformer2DModel(nn.Module):
             noise_rotary_emb, ref_img_rotary_emb,
             l_effective_ref_img_len, l_effective_img_len,
             temb,
+            transformer_options=transformer_options,
         )
 
         hidden_states = torch.cat([text_hidden_states, combined_img_hidden_states], dim=1)
         attention_mask = None
 
         for layer in self.layers:
-            hidden_states = layer(hidden_states, attention_mask, rotary_emb, temb)
+            hidden_states = layer(hidden_states, attention_mask, rotary_emb, temb, transformer_options=transformer_options)
 
         hidden_states = self.norm_out(hidden_states, temb)
 

comfy/ldm/qwen_image/controlnet.py

@@ -0,0 +1,77 @@
+import torch
+import math
+
+from .model import QwenImageTransformer2DModel
+
+
+class QwenImageControlNetModel(QwenImageTransformer2DModel):
+    def __init__(
+        self,
+        extra_condition_channels=0,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs
+    ):
+        super().__init__(final_layer=False, dtype=dtype, device=device, operations=operations, **kwargs)
+        self.main_model_double = 60
+
+        # controlnet_blocks
+        self.controlnet_blocks = torch.nn.ModuleList([])
+        for _ in range(len(self.transformer_blocks)):
+            self.controlnet_blocks.append(operations.Linear(self.inner_dim, self.inner_dim, device=device, dtype=dtype))
+        self.controlnet_x_embedder = operations.Linear(self.in_channels + extra_condition_channels, self.inner_dim, device=device, dtype=dtype)
+
+    def forward(
+        self,
+        x,
+        timesteps,
+        context,
+        attention_mask=None,
+        guidance: torch.Tensor = None,
+        ref_latents=None,
+        hint=None,
+        transformer_options={},
+        **kwargs
+    ):
+        timestep = timesteps
+        encoder_hidden_states = context
+        encoder_hidden_states_mask = attention_mask
+
+        hidden_states, img_ids, orig_shape = self.process_img(x)
+        hint, _, _ = self.process_img(hint)
+
+        txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
+        txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        del ids, txt_ids, img_ids
+
+        hidden_states = self.img_in(hidden_states) + self.controlnet_x_embedder(hint)
+        encoder_hidden_states = self.txt_norm(encoder_hidden_states)
+        encoder_hidden_states = self.txt_in(encoder_hidden_states)
+
+        if guidance is not None:
+            guidance = guidance * 1000
+
+        temb = (
+            self.time_text_embed(timestep, hidden_states)
+            if guidance is None
+            else self.time_text_embed(timestep, guidance, hidden_states)
+        )
+
+        repeat = math.ceil(self.main_model_double / len(self.controlnet_blocks))
+
+        controlnet_block_samples = ()
+        for i, block in enumerate(self.transformer_blocks):
+            encoder_hidden_states, hidden_states = block(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_hidden_states_mask=encoder_hidden_states_mask,
+                temb=temb,
+                image_rotary_emb=image_rotary_emb,
+            )
+
+            controlnet_block_samples = controlnet_block_samples + (self.controlnet_blocks[i](hidden_states),) * repeat
+
+        return {"input": controlnet_block_samples[:self.main_model_double]}

comfy/ldm/qwen_image/model.py

@@ -9,6 +9,7 @@ from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps
 from comfy.ldm.modules.attention import optimized_attention_masked
 from comfy.ldm.flux.layers import EmbedND
 import comfy.ldm.common_dit
+import comfy.patcher_extension
 
 class GELU(nn.Module):
     def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True, dtype=None, device=None, operations=None):
@@ -131,6 +132,7 @@ class Attention(nn.Module):
         encoder_hidden_states_mask: torch.FloatTensor = None,
         attention_mask: Optional[torch.FloatTensor] = None,
         image_rotary_emb: Optional[torch.Tensor] = None,
+        transformer_options={},
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         seq_txt = encoder_hidden_states.shape[1]
 
@@ -158,7 +160,7 @@ class Attention(nn.Module):
         joint_key = joint_key.flatten(start_dim=2)
         joint_value = joint_value.flatten(start_dim=2)
 
-        joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads, attention_mask)
+        joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads, attention_mask, transformer_options=transformer_options)
 
         txt_attn_output = joint_hidden_states[:, :seq_txt, :]
         img_attn_output = joint_hidden_states[:, seq_txt:, :]
@@ -214,9 +216,9 @@ class QwenImageTransformerBlock(nn.Module):
             operations=operations,
         )
 
-    def _modulate(self, x, mod_params):
-        shift, scale, gate = mod_params.chunk(3, dim=-1)
-        return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1), gate.unsqueeze(1)
+    def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        shift, scale, gate = torch.chunk(mod_params, 3, dim=-1)
+        return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1)
 
     def forward(
         self,
@@ -225,6 +227,7 @@ class QwenImageTransformerBlock(nn.Module):
         encoder_hidden_states_mask: torch.Tensor,
         temb: torch.Tensor,
         image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        transformer_options={},
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         img_mod_params = self.img_mod(temb)
         txt_mod_params = self.txt_mod(temb)
@@ -241,6 +244,7 @@ class QwenImageTransformerBlock(nn.Module):
             encoder_hidden_states=txt_modulated,
             encoder_hidden_states_mask=encoder_hidden_states_mask,
             image_rotary_emb=image_rotary_emb,
+            transformer_options=transformer_options,
         )
 
         hidden_states = hidden_states + img_gate1 * img_attn_output
@@ -248,11 +252,11 @@ class QwenImageTransformerBlock(nn.Module):
 
         img_normed2 = self.img_norm2(hidden_states)
         img_modulated2, img_gate2 = self._modulate(img_normed2, img_mod2)
-        hidden_states = hidden_states + img_gate2 * self.img_mlp(img_modulated2)
+        hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2))
 
         txt_normed2 = self.txt_norm2(encoder_hidden_states)
         txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2)
-        encoder_hidden_states = encoder_hidden_states + txt_gate2 * self.txt_mlp(txt_modulated2)
+        encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2))
 
         return encoder_hidden_states, hidden_states
 
@@ -275,7 +279,7 @@ class LastLayer(nn.Module):
     def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:
         emb = self.linear(self.silu(conditioning_embedding))
         scale, shift = torch.chunk(emb, 2, dim=1)
-        x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
+        x = torch.addcmul(shift[:, None, :], self.norm(x), (1 + scale)[:, None, :])
         return x
 
 
@@ -293,6 +297,7 @@ class QwenImageTransformer2DModel(nn.Module):
         guidance_embeds: bool = False,
         axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
         image_model=None,
+        final_layer=True,
         dtype=None,
         device=None,
         operations=None,
@@ -300,6 +305,7 @@ class QwenImageTransformer2DModel(nn.Module):
         super().__init__()
         self.dtype = dtype
         self.patch_size = patch_size
+        self.in_channels = in_channels
         self.out_channels = out_channels or in_channels
         self.inner_dim = num_attention_heads * attention_head_dim
 
@@ -329,46 +335,86 @@ class QwenImageTransformer2DModel(nn.Module):
             for _ in range(num_layers)
         ])
 
-        self.norm_out = LastLayer(self.inner_dim, self.inner_dim, dtype=dtype, device=device, operations=operations)
-        self.proj_out = operations.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, dtype=dtype, device=device)
-        self.gradient_checkpointing = False
+        if final_layer:
+            self.norm_out = LastLayer(self.inner_dim, self.inner_dim, dtype=dtype, device=device, operations=operations)
+            self.proj_out = operations.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, dtype=dtype, device=device)
 
-    def pos_embeds(self, x, context):
+    def process_img(self, x, index=0, h_offset=0, w_offset=0):
         bs, c, t, h, w = x.shape
         patch_size = self.patch_size
+        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (1, self.patch_size, self.patch_size))
+        orig_shape = hidden_states.shape
+        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
+        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5)
+        hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
         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).unsqueeze(1)
-        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
-        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+        h_offset = ((h_offset + (patch_size // 2)) // patch_size)
+        w_offset = ((w_offset + (patch_size // 2)) // patch_size)
 
-        txt_start = round(max(h_len, w_len))
-        txt_ids = torch.linspace(txt_start, txt_start + context.shape[1], steps=context.shape[1], device=x.device, dtype=x.dtype).reshape(1, -1, 1).repeat(bs, 1, 3)
-        ids = torch.cat((txt_ids, img_ids), dim=1)
-        return self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device)
+        img_ids[:, :, 0] = img_ids[:, :, 1] + index
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - (h_len // 2)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) - (w_len // 2)
+        return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shape
 
-    def forward(
+    def forward(self, x, timestep, context, attention_mask=None, guidance=None, ref_latents=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, attention_mask, guidance, ref_latents, transformer_options, **kwargs)
+
+    def _forward(
         self,
         x,
         timesteps,
         context,
         attention_mask=None,
         guidance: torch.Tensor = None,
+        ref_latents=None,
+        transformer_options={},
+        control=None,
         **kwargs
     ):
         timestep = timesteps
         encoder_hidden_states = context
         encoder_hidden_states_mask = attention_mask
 
-        image_rotary_emb = self.pos_embeds(x, context)
+        hidden_states, img_ids, orig_shape = self.process_img(x)
+        num_embeds = hidden_states.shape[1]
 
-        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (1, self.patch_size, self.patch_size))
-        orig_shape = hidden_states.shape
-        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
-        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5)
-        hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
+        if ref_latents is not None:
+            h = 0
+            w = 0
+            index = 0
+            index_ref_method = kwargs.get("ref_latents_method", "index") == "index"
+            for ref in ref_latents:
+                if index_ref_method:
+                    index += 1
+                    h_offset = 0
+                    w_offset = 0
+                else:
+                    index = 1
+                    h_offset = 0
+                    w_offset = 0
+                    if ref.shape[-2] + h > ref.shape[-1] + w:
+                        w_offset = w
+                    else:
+                        h_offset = h
+                    h = max(h, ref.shape[-2] + h_offset)
+                    w = max(w, ref.shape[-1] + w_offset)
+
+                kontext, kontext_ids, _ = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
+                hidden_states = torch.cat([hidden_states, kontext], dim=1)
+                img_ids = torch.cat([img_ids, kontext_ids], dim=1)
+
+        txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
+        txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        del ids, txt_ids, img_ids
 
         hidden_states = self.img_in(hidden_states)
         encoder_hidden_states = self.txt_norm(encoder_hidden_states)
@@ -383,18 +429,45 @@ class QwenImageTransformer2DModel(nn.Module):
             else self.time_text_embed(timestep, guidance, hidden_states)
         )
 
-        for block in self.transformer_blocks:
-            encoder_hidden_states, hidden_states = block(
-                hidden_states=hidden_states,
-                encoder_hidden_states=encoder_hidden_states,
-                encoder_hidden_states_mask=encoder_hidden_states_mask,
-                temb=temb,
-                image_rotary_emb=image_rotary_emb,
-            )
+        patches_replace = transformer_options.get("patches_replace", {})
+        patches = transformer_options.get("patches", {})
+        blocks_replace = patches_replace.get("dit", {})
+
+        for i, block in enumerate(self.transformer_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], transformer_options=args["transformer_options"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": hidden_states, "txt": encoder_hidden_states, "vec": temb, "pe": image_rotary_emb, "transformer_options": transformer_options}, {"original_block": block_wrap})
+                hidden_states = out["img"]
+                encoder_hidden_states = out["txt"]
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_hidden_states_mask=encoder_hidden_states_mask,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                    transformer_options=transformer_options,
+                )
+
+            if "double_block" in patches:
+                for p in patches["double_block"]:
+                    out = p({"img": hidden_states, "txt": encoder_hidden_states, "x": x, "block_index": i, "transformer_options": transformer_options})
+                    hidden_states = out["img"]
+                    encoder_hidden_states = out["txt"]
+
+            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:
+                        hidden_states[:, :add.shape[1]] += add
 
         hidden_states = self.norm_out(hidden_states, temb)
         hidden_states = self.proj_out(hidden_states)
 
-        hidden_states = hidden_states.view(orig_shape[0], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
+        hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
         hidden_states = hidden_states.permute(0, 3, 1, 4, 2, 5)
         return hidden_states.reshape(orig_shape)[:, :, :, :x.shape[-2], :x.shape[-1]]

comfy/ldm/wan/model.py

@@ -4,13 +4,14 @@ import math
 
 import torch
 import torch.nn as nn
-from einops import repeat
+from einops import rearrange
 
 from comfy.ldm.modules.attention import optimized_attention
 from comfy.ldm.flux.layers import EmbedND
 from comfy.ldm.flux.math import apply_rope
 import comfy.ldm.common_dit
 import comfy.model_management
+import comfy.patcher_extension
 
 
 def sinusoidal_embedding_1d(dim, position):
@@ -51,7 +52,7 @@ class WanSelfAttention(nn.Module):
         self.norm_q = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
         self.norm_k = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
 
-    def forward(self, x, freqs):
+    def forward(self, x, freqs, transformer_options={}):
         r"""
         Args:
             x(Tensor): Shape [B, L, num_heads, C / num_heads]
@@ -74,6 +75,7 @@ class WanSelfAttention(nn.Module):
             k.view(b, s, n * d),
             v,
             heads=self.num_heads,
+            transformer_options=transformer_options,
         )
 
         x = self.o(x)
@@ -82,7 +84,7 @@ class WanSelfAttention(nn.Module):
 
 class WanT2VCrossAttention(WanSelfAttention):
 
-    def forward(self, x, context, **kwargs):
+    def forward(self, x, context, transformer_options={}, **kwargs):
         r"""
         Args:
             x(Tensor): Shape [B, L1, C]
@@ -94,7 +96,7 @@ class WanT2VCrossAttention(WanSelfAttention):
         v = self.v(context)
 
         # compute attention
-        x = optimized_attention(q, k, v, heads=self.num_heads)
+        x = optimized_attention(q, k, v, heads=self.num_heads, transformer_options=transformer_options)
 
         x = self.o(x)
         return x
@@ -115,7 +117,7 @@ class WanI2VCrossAttention(WanSelfAttention):
         # self.alpha = nn.Parameter(torch.zeros((1, )))
         self.norm_k_img = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
 
-    def forward(self, x, context, context_img_len):
+    def forward(self, x, context, context_img_len, transformer_options={}):
         r"""
         Args:
             x(Tensor): Shape [B, L1, C]
@@ -130,9 +132,9 @@ class WanI2VCrossAttention(WanSelfAttention):
         v = self.v(context)
         k_img = self.norm_k_img(self.k_img(context_img))
         v_img = self.v_img(context_img)
-        img_x = optimized_attention(q, k_img, v_img, heads=self.num_heads)
+        img_x = optimized_attention(q, k_img, v_img, heads=self.num_heads, transformer_options=transformer_options)
         # compute attention
-        x = optimized_attention(q, k, v, heads=self.num_heads)
+        x = optimized_attention(q, k, v, heads=self.num_heads, transformer_options=transformer_options)
 
         # output
         x = x + img_x
@@ -148,11 +150,14 @@ WAN_CROSSATTENTION_CLASSES = {
 
 def repeat_e(e, x):
     repeats = 1
-    if e.shape[1] > 1:
-        repeats = x.shape[1] // e.shape[1]
+    if e.size(1) > 1:
+        repeats = x.size(1) // e.size(1)
     if repeats == 1:
         return e
-    return torch.repeat_interleave(e, repeats, dim=1)
+    if repeats * e.size(1) == x.size(1):
+        return torch.repeat_interleave(e, repeats, dim=1)
+    else:
+        return torch.repeat_interleave(e, repeats + 1, dim=1)[:, :x.size(1)]
 
 
 class WanAttentionBlock(nn.Module):
@@ -202,6 +207,7 @@ class WanAttentionBlock(nn.Module):
         freqs,
         context,
         context_img_len=257,
+        transformer_options={},
     ):
         r"""
         Args:
@@ -219,15 +225,15 @@ class WanAttentionBlock(nn.Module):
 
         # self-attention
         y = self.self_attn(
-            self.norm1(x) * (1 + repeat_e(e[1], x)) + repeat_e(e[0], x),
-            freqs)
+            torch.addcmul(repeat_e(e[0], x), self.norm1(x), 1 + repeat_e(e[1], x)),
+            freqs, transformer_options=transformer_options)
 
-        x = x + y * repeat_e(e[2], x)
+        x = torch.addcmul(x, y, repeat_e(e[2], x))
 
         # cross-attention & ffn
-        x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len)
-        y = self.ffn(self.norm2(x) * (1 + repeat_e(e[4], x)) + repeat_e(e[3], x))
-        x = x + y * repeat_e(e[5], x)
+        x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len, transformer_options=transformer_options)
+        y = self.ffn(torch.addcmul(repeat_e(e[3], x), self.norm2(x), 1 + repeat_e(e[4], x)))
+        x = torch.addcmul(x, y, repeat_e(e[5], x))
         return x
 
 
@@ -342,7 +348,7 @@ class Head(nn.Module):
         else:
             e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e.unsqueeze(2)).unbind(2)
 
-        x = (self.head(self.norm(x) * (1 + repeat_e(e[1], x)) + repeat_e(e[0], x)))
+        x = (self.head(torch.addcmul(repeat_e(e[0], x), self.norm(x), 1 + repeat_e(e[1], x))))
         return x
 
 
@@ -391,6 +397,7 @@ class WanModel(torch.nn.Module):
                  cross_attn_norm=True,
                  eps=1e-6,
                  flf_pos_embed_token_number=None,
+                 in_dim_ref_conv=None,
                  image_model=None,
                  device=None,
                  dtype=None,
@@ -484,6 +491,11 @@ class WanModel(torch.nn.Module):
         else:
             self.img_emb = None
 
+        if in_dim_ref_conv is not None:
+            self.ref_conv = operations.Conv2d(in_dim_ref_conv, dim, kernel_size=patch_size[1:], stride=patch_size[1:], device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        else:
+            self.ref_conv = None
+
     def forward_orig(
         self,
         x,
@@ -526,6 +538,13 @@ class WanModel(torch.nn.Module):
         e = e.reshape(t.shape[0], -1, e.shape[-1])
         e0 = self.time_projection(e).unflatten(2, (6, self.dim))
 
+        full_ref = None
+        if self.ref_conv is not None:
+            full_ref = kwargs.get("reference_latent", None)
+            if full_ref is not None:
+                full_ref = self.ref_conv(full_ref).flatten(2).transpose(1, 2)
+                x = torch.concat((full_ref, x), dim=1)
+
         # context
         context = self.text_embedding(context)
 
@@ -542,41 +561,66 @@ class WanModel(torch.nn.Module):
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len, transformer_options=args["transformer_options"])
                     return out
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 x = out["img"]
             else:
-                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len, transformer_options=transformer_options)
 
         # head
         x = self.head(x, e)
 
+        if full_ref is not None:
+            x = x[:, full_ref.shape[1]:]
+
         # unpatchify
         x = self.unpatchify(x, grid_sizes)
         return x
 
-    def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
-        bs, c, t, h, w = x.shape
-        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
-
+    def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None):
         patch_size = self.patch_size
         t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
         h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
         w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
 
+        if steps_t is None:
+            steps_t = t_len
+        if steps_h is None:
+            steps_h = h_len
+        if steps_w is None:
+            steps_w = w_len
+
+        img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype)
+        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1)
+        img_ids = img_ids.reshape(1, -1, img_ids.shape[-1])
+
+        freqs = self.rope_embedder(img_ids).movedim(1, 2)
+        return freqs
+
+    def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, clip_fea, time_dim_concat, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
+        bs, c, t, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+
+        t_len = t
         if time_dim_concat is not None:
             time_dim_concat = comfy.ldm.common_dit.pad_to_patch_size(time_dim_concat, self.patch_size)
             x = torch.cat([x, time_dim_concat], dim=2)
-            t_len = ((x.shape[2] + (patch_size[0] // 2)) // patch_size[0])
+            t_len = x.shape[2]
 
-        img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device, dtype=x.dtype)
-        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1)
-        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1)
-        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1)
-        img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+        if self.ref_conv is not None and "reference_latent" in kwargs:
+            t_len += 1
 
-        freqs = self.rope_embedder(img_ids).movedim(1, 2)
+        freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype)
         return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w]
 
     def unpatchify(self, x, grid_sizes):
@@ -700,17 +744,17 @@ class VaceWanModel(WanModel):
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len, transformer_options=args["transformer_options"])
                     return out
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 x = out["img"]
             else:
-                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len, transformer_options=transformer_options)
 
             ii = self.vace_layers_mapping.get(i, None)
             if ii is not None:
                 for iii in range(len(c)):
-                    c_skip, c[iii] = self.vace_blocks[ii](c[iii], x=x_orig, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+                    c_skip, c[iii] = self.vace_blocks[ii](c[iii], x=x_orig, e=e0, freqs=freqs, context=context, context_img_len=context_img_len, transformer_options=transformer_options)
                     x += c_skip * vace_strength[iii]
                 del c_skip
         # head
@@ -749,7 +793,12 @@ class CameraWanModel(WanModel):
                  operations=None,
                  ):
 
-        super().__init__(model_type='i2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
+        if model_type == 'camera':
+            model_type = 'i2v'
+        else:
+            model_type = 't2v'
+
+        super().__init__(model_type=model_type, patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
         operation_settings = {"operations": operations, "device": device, "dtype": dtype}
 
         self.control_adapter = WanCamAdapter(in_dim_control_adapter, dim, kernel_size=patch_size[1:], stride=patch_size[1:], operation_settings=operation_settings)
@@ -794,12 +843,12 @@ class CameraWanModel(WanModel):
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len, transformer_options=args["transformer_options"])
                     return out
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 x = out["img"]
             else:
-                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len, transformer_options=transformer_options)
 
         # head
         x = self.head(x, e)
@@ -807,3 +856,468 @@ class CameraWanModel(WanModel):
         # unpatchify
         x = self.unpatchify(x, grid_sizes)
         return x
+
+
+class CausalConv1d(nn.Module):
+
+    def __init__(self,
+                 chan_in,
+                 chan_out,
+                 kernel_size=3,
+                 stride=1,
+                 dilation=1,
+                 pad_mode='replicate',
+                 operations=None,
+                 **kwargs):
+        super().__init__()
+
+        self.pad_mode = pad_mode
+        padding = (kernel_size - 1, 0)  # T
+        self.time_causal_padding = padding
+
+        self.conv = operations.Conv1d(
+            chan_in,
+            chan_out,
+            kernel_size,
+            stride=stride,
+            dilation=dilation,
+            **kwargs)
+
+    def forward(self, x):
+        x = torch.nn.functional.pad(x, self.time_causal_padding, mode=self.pad_mode)
+        return self.conv(x)
+
+
+class MotionEncoder_tc(nn.Module):
+
+    def __init__(self,
+                 in_dim: int,
+                 hidden_dim: int,
+                 num_heads=int,
+                 need_global=True,
+                 dtype=None,
+                 device=None,
+                 operations=None,):
+        factory_kwargs = {"dtype": dtype, "device": device}
+        super().__init__()
+
+        self.num_heads = num_heads
+        self.need_global = need_global
+        self.conv1_local = CausalConv1d(in_dim, hidden_dim // 4 * num_heads, 3, stride=1, operations=operations, **factory_kwargs)
+        if need_global:
+            self.conv1_global = CausalConv1d(
+                in_dim, hidden_dim // 4, 3, stride=1, operations=operations, **factory_kwargs)
+        self.norm1 = operations.LayerNorm(
+            hidden_dim // 4,
+            elementwise_affine=False,
+            eps=1e-6,
+            **factory_kwargs)
+        self.act = nn.SiLU()
+        self.conv2 = CausalConv1d(hidden_dim // 4, hidden_dim // 2, 3, stride=2, operations=operations, **factory_kwargs)
+        self.conv3 = CausalConv1d(hidden_dim // 2, hidden_dim, 3, stride=2, operations=operations, **factory_kwargs)
+
+        if need_global:
+            self.final_linear = operations.Linear(hidden_dim, hidden_dim, **factory_kwargs)
+
+        self.norm1 = operations.LayerNorm(
+            hidden_dim // 4,
+            elementwise_affine=False,
+            eps=1e-6,
+            **factory_kwargs)
+
+        self.norm2 = operations.LayerNorm(
+            hidden_dim // 2,
+            elementwise_affine=False,
+            eps=1e-6,
+            **factory_kwargs)
+
+        self.norm3 = operations.LayerNorm(
+            hidden_dim, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+        self.padding_tokens = nn.Parameter(torch.empty(1, 1, 1, hidden_dim, **factory_kwargs))
+
+    def forward(self, x):
+        x = rearrange(x, 'b t c -> b c t')
+        x_ori = x.clone()
+        b, c, t = x.shape
+        x = self.conv1_local(x)
+        x = rearrange(x, 'b (n c) t -> (b n) t c', n=self.num_heads)
+        x = self.norm1(x)
+        x = self.act(x)
+        x = rearrange(x, 'b t c -> b c t')
+        x = self.conv2(x)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm2(x)
+        x = self.act(x)
+        x = rearrange(x, 'b t c -> b c t')
+        x = self.conv3(x)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm3(x)
+        x = self.act(x)
+        x = rearrange(x, '(b n) t c -> b t n c', b=b)
+        padding = comfy.model_management.cast_to(self.padding_tokens, dtype=x.dtype, device=x.device).repeat(b, x.shape[1], 1, 1)
+        x = torch.cat([x, padding], dim=-2)
+        x_local = x.clone()
+
+        if not self.need_global:
+            return x_local
+
+        x = self.conv1_global(x_ori)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm1(x)
+        x = self.act(x)
+        x = rearrange(x, 'b t c -> b c t')
+        x = self.conv2(x)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm2(x)
+        x = self.act(x)
+        x = rearrange(x, 'b t c -> b c t')
+        x = self.conv3(x)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm3(x)
+        x = self.act(x)
+        x = self.final_linear(x)
+        x = rearrange(x, '(b n) t c -> b t n c', b=b)
+
+        return x, x_local
+
+
+class CausalAudioEncoder(nn.Module):
+
+    def __init__(self,
+                 dim=5120,
+                 num_layers=25,
+                 out_dim=2048,
+                 video_rate=8,
+                 num_token=4,
+                 need_global=False,
+                 dtype=None,
+                 device=None,
+                 operations=None):
+        super().__init__()
+        self.encoder = MotionEncoder_tc(
+            in_dim=dim,
+            hidden_dim=out_dim,
+            num_heads=num_token,
+            need_global=need_global, dtype=dtype, device=device, operations=operations)
+        weight = torch.empty((1, num_layers, 1, 1), dtype=dtype, device=device)
+
+        self.weights = torch.nn.Parameter(weight)
+        self.act = torch.nn.SiLU()
+
+    def forward(self, features):
+        # features B * num_layers * dim * video_length
+        weights = self.act(comfy.model_management.cast_to(self.weights, dtype=features.dtype, device=features.device))
+        weights_sum = weights.sum(dim=1, keepdims=True)
+        weighted_feat = ((features * weights) / weights_sum).sum(
+            dim=1)  # b dim f
+        weighted_feat = weighted_feat.permute(0, 2, 1)  # b f dim
+        res = self.encoder(weighted_feat)  # b f n dim
+        return res  # b f n dim
+
+
+class AdaLayerNorm(nn.Module):
+    def __init__(self, embedding_dim, output_dim=None, norm_elementwise_affine=False, norm_eps=1e-5, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        output_dim = output_dim or embedding_dim * 2
+
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(embedding_dim, output_dim, dtype=dtype, device=device)
+        self.norm = operations.LayerNorm(output_dim // 2, norm_eps, norm_elementwise_affine, dtype=dtype, device=device)
+
+    def forward(self, x, temb):
+        temb = self.linear(self.silu(temb))
+        shift, scale = temb.chunk(2, dim=1)
+        shift = shift[:, None, :]
+        scale = scale[:, None, :]
+        x = self.norm(x) * (1 + scale) + shift
+        return x
+
+
+class AudioInjector_WAN(nn.Module):
+
+    def __init__(self,
+                 dim=2048,
+                 num_heads=32,
+                 inject_layer=[0, 27],
+                 root_net=None,
+                 enable_adain=False,
+                 adain_dim=2048,
+                 adain_mode=None,
+                 dtype=None,
+                 device=None,
+                 operations=None):
+        super().__init__()
+        self.enable_adain = enable_adain
+        self.adain_mode = adain_mode
+        self.injected_block_id = {}
+        audio_injector_id = 0
+        for inject_id in inject_layer:
+            self.injected_block_id[inject_id] = audio_injector_id
+            audio_injector_id += 1
+
+        self.injector = nn.ModuleList([
+            WanT2VCrossAttention(
+                dim=dim,
+                num_heads=num_heads,
+                qk_norm=True, operation_settings={"operations": operations, "device": device, "dtype": dtype}
+            ) for _ in range(audio_injector_id)
+        ])
+        self.injector_pre_norm_feat = nn.ModuleList([
+            operations.LayerNorm(
+                dim,
+                elementwise_affine=False,
+                eps=1e-6, dtype=dtype, device=device
+            ) for _ in range(audio_injector_id)
+        ])
+        self.injector_pre_norm_vec = nn.ModuleList([
+            operations.LayerNorm(
+                dim,
+                elementwise_affine=False,
+                eps=1e-6, dtype=dtype, device=device
+            ) for _ in range(audio_injector_id)
+        ])
+        if enable_adain:
+            self.injector_adain_layers = nn.ModuleList([
+                AdaLayerNorm(
+                    output_dim=dim * 2, embedding_dim=adain_dim, dtype=dtype, device=device, operations=operations)
+                for _ in range(audio_injector_id)
+            ])
+            if adain_mode != "attn_norm":
+                self.injector_adain_output_layers = nn.ModuleList(
+                    [operations.Linear(dim, dim, dtype=dtype, device=device) for _ in range(audio_injector_id)])
+
+    def forward(self, x, block_id, audio_emb, audio_emb_global, seq_len):
+        audio_attn_id = self.injected_block_id.get(block_id, None)
+        if audio_attn_id is None:
+            return x
+
+        num_frames = audio_emb.shape[1]
+        input_hidden_states = rearrange(x[:, :seq_len], "b (t n) c -> (b t) n c", t=num_frames)
+        if self.enable_adain and self.adain_mode == "attn_norm":
+            audio_emb_global = rearrange(audio_emb_global, "b t n c -> (b t) n c")
+            adain_hidden_states = self.injector_adain_layers[audio_attn_id](input_hidden_states, temb=audio_emb_global[:, 0])
+            attn_hidden_states = adain_hidden_states
+        else:
+            attn_hidden_states = self.injector_pre_norm_feat[audio_attn_id](input_hidden_states)
+        audio_emb = rearrange(audio_emb, "b t n c -> (b t) n c", t=num_frames)
+        attn_audio_emb = audio_emb
+        residual_out = self.injector[audio_attn_id](x=attn_hidden_states, context=attn_audio_emb)
+        residual_out = rearrange(
+            residual_out, "(b t) n c -> b (t n) c", t=num_frames)
+        x[:, :seq_len] = x[:, :seq_len] + residual_out
+        return x
+
+
+class FramePackMotioner(nn.Module):
+    def __init__(
+            self,
+            inner_dim=1024,
+            num_heads=16,  # Used to indicate the number of heads in the backbone network; unrelated to this module's design
+            zip_frame_buckets=[
+                1, 2, 16
+            ],  # Three numbers representing the number of frames sampled for patch operations from the nearest to the farthest frames
+            drop_mode="drop",  # If not "drop", it will use "padd", meaning padding instead of deletion
+            dtype=None,
+            device=None,
+            operations=None):
+        super().__init__()
+        self.proj = operations.Conv3d(16, inner_dim, kernel_size=(1, 2, 2), stride=(1, 2, 2), dtype=dtype, device=device)
+        self.proj_2x = operations.Conv3d(16, inner_dim, kernel_size=(2, 4, 4), stride=(2, 4, 4), dtype=dtype, device=device)
+        self.proj_4x = operations.Conv3d(16, inner_dim, kernel_size=(4, 8, 8), stride=(4, 8, 8), dtype=dtype, device=device)
+        self.zip_frame_buckets = zip_frame_buckets
+
+        self.inner_dim = inner_dim
+        self.num_heads = num_heads
+
+        self.drop_mode = drop_mode
+
+    def forward(self, motion_latents, rope_embedder, add_last_motion=2):
+        lat_height, lat_width = motion_latents.shape[3], motion_latents.shape[4]
+        padd_lat = torch.zeros(motion_latents.shape[0], 16, sum(self.zip_frame_buckets), lat_height, lat_width).to(device=motion_latents.device, dtype=motion_latents.dtype)
+        overlap_frame = min(padd_lat.shape[2], motion_latents.shape[2])
+        if overlap_frame > 0:
+            padd_lat[:, :, -overlap_frame:] = motion_latents[:, :, -overlap_frame:]
+
+        if add_last_motion < 2 and self.drop_mode != "drop":
+            zero_end_frame = sum(self.zip_frame_buckets[:len(self.zip_frame_buckets) - add_last_motion - 1])
+            padd_lat[:, :, -zero_end_frame:] = 0
+
+        clean_latents_4x, clean_latents_2x, clean_latents_post = padd_lat[:, :, -sum(self.zip_frame_buckets):, :, :].split(self.zip_frame_buckets[::-1], dim=2)  # 16, 2 ,1
+
+        # patchfy
+        clean_latents_post = self.proj(clean_latents_post).flatten(2).transpose(1, 2)
+        clean_latents_2x = self.proj_2x(clean_latents_2x)
+        l_2x_shape = clean_latents_2x.shape
+        clean_latents_2x = clean_latents_2x.flatten(2).transpose(1, 2)
+        clean_latents_4x = self.proj_4x(clean_latents_4x)
+        l_4x_shape = clean_latents_4x.shape
+        clean_latents_4x = clean_latents_4x.flatten(2).transpose(1, 2)
+
+        if add_last_motion < 2 and self.drop_mode == "drop":
+            clean_latents_post = clean_latents_post[:, :
+                                                    0] if add_last_motion < 2 else clean_latents_post
+            clean_latents_2x = clean_latents_2x[:, :
+                                                0] if add_last_motion < 1 else clean_latents_2x
+
+        motion_lat = torch.cat([clean_latents_post, clean_latents_2x, clean_latents_4x], dim=1)
+
+        rope_post = rope_embedder.rope_encode(1, lat_height, lat_width, t_start=-1, device=motion_latents.device, dtype=motion_latents.dtype)
+        rope_2x = rope_embedder.rope_encode(1, lat_height, lat_width, t_start=-3, steps_h=l_2x_shape[-2], steps_w=l_2x_shape[-1], device=motion_latents.device, dtype=motion_latents.dtype)
+        rope_4x = rope_embedder.rope_encode(4, lat_height, lat_width, t_start=-19, steps_h=l_4x_shape[-2], steps_w=l_4x_shape[-1], device=motion_latents.device, dtype=motion_latents.dtype)
+
+        rope = torch.cat([rope_post, rope_2x, rope_4x], dim=1)
+        return motion_lat, rope
+
+
+class WanModel_S2V(WanModel):
+    def __init__(self,
+                 model_type='s2v',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 audio_dim=1024,
+                 num_audio_token=4,
+                 enable_adain=True,
+                 cond_dim=16,
+                 audio_inject_layers=[0, 4, 8, 12, 16, 20, 24, 27, 30, 33, 36, 39],
+                 adain_mode="attn_norm",
+                 framepack_drop_mode="padd",
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+
+        super().__init__(model_type='t2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, image_model=image_model, device=device, dtype=dtype, operations=operations)
+
+        self.trainable_cond_mask = operations.Embedding(3, self.dim, device=device, dtype=dtype)
+
+        self.casual_audio_encoder = CausalAudioEncoder(
+            dim=audio_dim,
+            out_dim=self.dim,
+            num_token=num_audio_token,
+            need_global=enable_adain, dtype=dtype, device=device, operations=operations)
+
+        if cond_dim > 0:
+            self.cond_encoder = operations.Conv3d(
+                cond_dim,
+                self.dim,
+                kernel_size=self.patch_size,
+                stride=self.patch_size, device=device, dtype=dtype)
+
+        self.audio_injector = AudioInjector_WAN(
+            dim=self.dim,
+            num_heads=self.num_heads,
+            inject_layer=audio_inject_layers,
+            root_net=self,
+            enable_adain=enable_adain,
+            adain_dim=self.dim,
+            adain_mode=adain_mode,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        self.frame_packer = FramePackMotioner(
+            inner_dim=self.dim,
+            num_heads=self.num_heads,
+            zip_frame_buckets=[1, 2, 16],
+            drop_mode=framepack_drop_mode,
+            dtype=dtype, device=device, operations=operations)
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        audio_embed=None,
+        reference_latent=None,
+        control_video=None,
+        reference_motion=None,
+        clip_fea=None,
+        freqs=None,
+        transformer_options={},
+        **kwargs,
+    ):
+        if audio_embed is not None:
+            num_embeds = x.shape[-3] * 4
+            audio_emb_global, audio_emb = self.casual_audio_encoder(audio_embed[:, :, :, :num_embeds])
+        else:
+            audio_emb = None
+
+        # embeddings
+        bs, _, time, height, width = x.shape
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        if control_video is not None:
+            x = x + self.cond_encoder(control_video)
+
+        if t.ndim == 1:
+            t = t.unsqueeze(1).repeat(1, x.shape[2])
+
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+        seq_len = x.size(1)
+
+        cond_mask_weight = comfy.model_management.cast_to(self.trainable_cond_mask.weight, dtype=x.dtype, device=x.device).unsqueeze(1).unsqueeze(1)
+        x = x + cond_mask_weight[0]
+
+        if reference_latent is not None:
+            ref = self.patch_embedding(reference_latent.float()).to(x.dtype)
+            ref = ref.flatten(2).transpose(1, 2)
+            freqs_ref = self.rope_encode(reference_latent.shape[-3], reference_latent.shape[-2], reference_latent.shape[-1], t_start=max(30, time + 9), device=x.device, dtype=x.dtype)
+            ref = ref + cond_mask_weight[1]
+            x = torch.cat([x, ref], dim=1)
+            freqs = torch.cat([freqs, freqs_ref], dim=1)
+            t = torch.cat([t, torch.zeros((t.shape[0], reference_latent.shape[-3]), device=t.device, dtype=t.dtype)], dim=1)
+            del ref, freqs_ref
+
+        if reference_motion is not None:
+            motion_encoded, freqs_motion = self.frame_packer(reference_motion, self)
+            motion_encoded = motion_encoded + cond_mask_weight[2]
+            x = torch.cat([x, motion_encoded], dim=1)
+            freqs = torch.cat([freqs, freqs_motion], dim=1)
+
+            t = torch.repeat_interleave(t, 2, dim=1)
+            t = torch.cat([t, torch.zeros((t.shape[0], 3), device=t.device, dtype=t.dtype)], dim=1)
+            del motion_encoded, freqs_motion
+
+        # time embeddings
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(dtype=x[0].dtype))
+        e = e.reshape(t.shape[0], -1, e.shape[-1])
+        e0 = self.time_projection(e).unflatten(2, (6, self.dim))
+
+        # context
+        context = self.text_embedding(context)
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context)
+            if audio_emb is not None:
+                x = self.audio_injector(x, i, audio_emb, audio_emb_global, seq_len)
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x

comfy/model_base.py

@@ -150,6 +150,7 @@ class BaseModel(torch.nn.Module):
         logging.debug("adm {}".format(self.adm_channels))
         self.memory_usage_factor = model_config.memory_usage_factor
         self.memory_usage_factor_conds = ()
+        self.memory_usage_shape_process = {}
 
     def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
@@ -350,8 +351,15 @@ class BaseModel(torch.nn.Module):
         input_shapes = [input_shape]
         for c in self.memory_usage_factor_conds:
             shape = cond_shapes.get(c, None)
-            if shape is not None and len(shape) > 0:
-                input_shapes += shape
+            if shape is not None:
+                if c in self.memory_usage_shape_process:
+                    out = []
+                    for s in shape:
+                        out.append(self.memory_usage_shape_process[c](s))
+                    shape = out
+
+                if len(shape) > 0:
+                    input_shapes += shape
 
         if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
             dtype = self.get_dtype()
@@ -890,6 +898,10 @@ class Flux(BaseModel):
             for lat in ref_latents:
                 latents.append(self.process_latent_in(lat))
             out['ref_latents'] = comfy.conds.CONDList(latents)
+
+            ref_latents_method = kwargs.get("reference_latents_method", None)
+            if ref_latents_method is not None:
+                out['ref_latents_method'] = comfy.conds.CONDConstant(ref_latents_method)
         return out
 
     def extra_conds_shapes(self, **kwargs):
@@ -1098,9 +1110,10 @@ class WAN21(BaseModel):
             shape_image[1] = extra_channels
             image = torch.zeros(shape_image, dtype=noise.dtype, layout=noise.layout, device=noise.device)
         else:
+            latent_dim = self.latent_format.latent_channels
             image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
-            for i in range(0, image.shape[1], 16):
-                image[:, i: i + 16] = self.process_latent_in(image[:, i: i + 16])
+            for i in range(0, image.shape[1], latent_dim):
+                image[:, i: i + latent_dim] = self.process_latent_in(image[:, i: i + latent_dim])
             image = utils.resize_to_batch_size(image, noise.shape[0])
 
         if extra_channels != image.shape[1] + 4:
@@ -1124,7 +1137,11 @@ class WAN21(BaseModel):
                 mask = mask.repeat(1, 4, 1, 1, 1)
             mask = utils.resize_to_batch_size(mask, noise.shape[0])
 
-        return torch.cat((mask, image), dim=1)
+        concat_mask_index = kwargs.get("concat_mask_index", 0)
+        if concat_mask_index != 0:
+            return torch.cat((image[:, :concat_mask_index], mask, image[:, concat_mask_index:]), dim=1)
+        else:
+            return torch.cat((mask, image), dim=1)
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
@@ -1140,6 +1157,10 @@ class WAN21(BaseModel):
         if time_dim_concat is not None:
             out['time_dim_concat'] = comfy.conds.CONDRegular(self.process_latent_in(time_dim_concat))
 
+        reference_latents = kwargs.get("reference_latents", None)
+        if reference_latents is not None:
+            out['reference_latent'] = comfy.conds.CONDRegular(self.process_latent_in(reference_latents[-1])[:, :, 0])
+
         return out
 
 
@@ -1189,18 +1210,50 @@ class WAN21_Camera(WAN21):
             out['camera_conditions'] = comfy.conds.CONDRegular(camera_conditions)
         return out
 
-class WAN22(BaseModel):
+class WAN22_S2V(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel_S2V)
+        self.memory_usage_factor_conds = ("reference_latent", "reference_motion")
+        self.memory_usage_shape_process = {"reference_motion": lambda shape: [shape[0], shape[1], 1.5, shape[-2], shape[-1]]}
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        audio_embed = kwargs.get("audio_embed", None)
+        if audio_embed is not None:
+            out['audio_embed'] = comfy.conds.CONDRegular(audio_embed)
+
+        reference_latents = kwargs.get("reference_latents", None)
+        if reference_latents is not None:
+            out['reference_latent'] = comfy.conds.CONDRegular(self.process_latent_in(reference_latents[-1]))
+
+        reference_motion = kwargs.get("reference_motion", None)
+        if reference_motion is not None:
+            out['reference_motion'] = comfy.conds.CONDRegular(self.process_latent_in(reference_motion))
+
+        control_video = kwargs.get("control_video", None)
+        if control_video is not None:
+            out['control_video'] = comfy.conds.CONDRegular(self.process_latent_in(control_video))
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['reference_latent'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
+
+        reference_motion = kwargs.get("reference_motion", None)
+        if reference_motion is not None:
+            out['reference_motion'] = reference_motion.shape
+        return out
+
+class WAN22(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
-        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
         self.image_to_video = image_to_video
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
-        cross_attn = kwargs.get("cross_attn", None)
-        if cross_attn is not None:
-            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
-
-        denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        denoise_mask = kwargs.get("denoise_mask", None)
         if denoise_mask is not None:
             out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
         return out
@@ -1313,10 +1366,28 @@ class Omnigen2(BaseModel):
 class QwenImage(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.qwen_image.model.QwenImageTransformer2DModel)
+        self.memory_usage_factor_conds = ("ref_latents",)
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
         cross_attn = kwargs.get("cross_attn", None)
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            latents = []
+            for lat in ref_latents:
+                latents.append(self.process_latent_in(lat))
+            out['ref_latents'] = comfy.conds.CONDList(latents)
+
+            ref_latents_method = kwargs.get("reference_latents_method", None)
+            if ref_latents_method is not None:
+                out['ref_latents_method'] = comfy.conds.CONDConstant(ref_latents_method)
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
         return out

comfy/model_detection.py

@@ -364,7 +364,12 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["vace_in_dim"] = state_dict['{}vace_patch_embedding.weight'.format(key_prefix)].shape[1]
             dit_config["vace_layers"] = count_blocks(state_dict_keys, '{}vace_blocks.'.format(key_prefix) + '{}.')
         elif '{}control_adapter.conv.weight'.format(key_prefix) in state_dict_keys:
-            dit_config["model_type"] = "camera"
+            if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
+                dit_config["model_type"] = "camera"
+            else:
+                dit_config["model_type"] = "camera_2.2"
+        elif '{}casual_audio_encoder.encoder.final_linear.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["model_type"] = "s2v"
         else:
             if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
                 dit_config["model_type"] = "i2v"
@@ -373,6 +378,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         flf_weight = state_dict.get('{}img_emb.emb_pos'.format(key_prefix))
         if flf_weight is not None:
             dit_config["flf_pos_embed_token_number"] = flf_weight.shape[1]
+
+        ref_conv_weight = state_dict.get('{}ref_conv.weight'.format(key_prefix))
+        if ref_conv_weight is not None:
+            dit_config["in_dim_ref_conv"] = ref_conv_weight.shape[1]
+
         return dit_config
 
     if '{}latent_in.weight'.format(key_prefix) in state_dict_keys:  # Hunyuan 3D
@@ -484,6 +494,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
     if '{}txt_norm.weight'.format(key_prefix) in state_dict_keys:  # Qwen Image
         dit_config = {}
         dit_config["image_model"] = "qwen_image"
+        dit_config["in_channels"] = state_dict['{}img_in.weight'.format(key_prefix)].shape[1]
+        dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
         return dit_config
 
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:

comfy/model_management.py

@@ -78,7 +78,6 @@ try:
     torch_version = torch.version.__version__
     temp = torch_version.split(".")
     torch_version_numeric = (int(temp[0]), int(temp[1]))
-    xpu_available = (torch_version_numeric[0] < 2 or (torch_version_numeric[0] == 2 and torch_version_numeric[1] <= 4)) and torch.xpu.is_available()
 except:
     pass
 
@@ -102,10 +101,14 @@ if args.directml is not None:
 
 try:
     import intel_extension_for_pytorch as ipex  # noqa: F401
-    _ = torch.xpu.device_count()
-    xpu_available = xpu_available or torch.xpu.is_available()
 except:
-    xpu_available = xpu_available or (hasattr(torch, "xpu") and torch.xpu.is_available())
+    pass
+
+try:
+    _ = torch.xpu.device_count()
+    xpu_available = torch.xpu.is_available()
+except:
+    xpu_available = False
 
 try:
     if torch.backends.mps.is_available():
@@ -590,7 +593,13 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
     else:
         minimum_memory_required = max(inference_memory, minimum_memory_required + extra_reserved_memory())
 
-    models = set(models)
+    models_temp = set()
+    for m in models:
+        models_temp.add(m)
+        for mm in m.model_patches_models():
+            models_temp.add(mm)
+
+    models = models_temp
 
     models_to_load = []
 
@@ -946,10 +955,12 @@ def pick_weight_dtype(dtype, fallback_dtype, device=None):
     return dtype
 
 def device_supports_non_blocking(device):
+    if args.force_non_blocking:
+        return True
     if is_device_mps(device):
         return False #pytorch bug? mps doesn't support non blocking
-    if is_intel_xpu():
-        return True
+    if is_intel_xpu(): #xpu does support non blocking but it is slower on iGPUs for some reason so disable by default until situation changes
+        return False
     if args.deterministic: #TODO: figure out why deterministic breaks non blocking from gpu to cpu (previews)
         return False
     if directml_enabled:
@@ -1282,10 +1293,10 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        if torch_version_numeric < (2, 6):
+        if torch_version_numeric < (2, 3):
             return True
         else:
-            return torch.xpu.get_device_capability(device)['has_bfloat16_conversions']
+            return torch.xpu.is_bf16_supported()
 
     if is_ascend_npu():
         return True

comfy/model_patcher.py

@@ -430,6 +430,9 @@ class ModelPatcher:
     def set_model_forward_timestep_embed_patch(self, patch):
         self.set_model_patch(patch, "forward_timestep_embed_patch")
 
+    def set_model_double_block_patch(self, patch):
+        self.set_model_patch(patch, "double_block")
+
     def add_object_patch(self, name, obj):
         self.object_patches[name] = obj
 
@@ -486,6 +489,30 @@ class ModelPatcher:
             if hasattr(wrap_func, "to"):
                 self.model_options["model_function_wrapper"] = wrap_func.to(device)
 
+    def model_patches_models(self):
+        to = self.model_options["transformer_options"]
+        models = []
+        if "patches" in to:
+            patches = to["patches"]
+            for name in patches:
+                patch_list = patches[name]
+                for i in range(len(patch_list)):
+                    if hasattr(patch_list[i], "models"):
+                        models += patch_list[i].models()
+        if "patches_replace" in to:
+            patches = to["patches_replace"]
+            for name in patches:
+                patch_list = patches[name]
+                for k in patch_list:
+                    if hasattr(patch_list[k], "models"):
+                        models += patch_list[k].models()
+        if "model_function_wrapper" in self.model_options:
+            wrap_func = self.model_options["model_function_wrapper"]
+            if hasattr(wrap_func, "models"):
+                models += wrap_func.models()
+
+        return models
+
     def model_dtype(self):
         if hasattr(self.model, "get_dtype"):
             return self.model.get_dtype()

comfy/ops.py

@@ -24,6 +24,32 @@ import comfy.float
 import comfy.rmsnorm
 import contextlib
 
+
+def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+    return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
+
+
+try:
+    if torch.cuda.is_available():
+        from torch.nn.attention import SDPBackend, sdpa_kernel
+        import inspect
+        if "set_priority" in inspect.signature(sdpa_kernel).parameters:
+            SDPA_BACKEND_PRIORITY = [
+                SDPBackend.FLASH_ATTENTION,
+                SDPBackend.EFFICIENT_ATTENTION,
+                SDPBackend.MATH,
+            ]
+
+            SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
+
+            def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+                with sdpa_kernel(SDPA_BACKEND_PRIORITY, set_priority=True):
+                    return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
+        else:
+            logging.warning("Torch version too old to set sdpa backend priority.")
+except (ModuleNotFoundError, TypeError):
+    logging.warning("Could not set sdpa backend priority.")
+
 cast_to = comfy.model_management.cast_to #TODO: remove once no more references
 
 def cast_to_input(weight, input, non_blocking=False, copy=True):

comfy/patcher_extension.py

@@ -50,6 +50,7 @@ class WrappersMP:
     OUTER_SAMPLE = "outer_sample"
     PREPARE_SAMPLING = "prepare_sampling"
     SAMPLER_SAMPLE = "sampler_sample"
+    PREDICT_NOISE = "predict_noise"
     CALC_COND_BATCH = "calc_cond_batch"
     APPLY_MODEL = "apply_model"
     DIFFUSION_MODEL = "diffusion_model"

comfy/rmsnorm.py

@@ -1,6 +1,7 @@
 import torch
 import comfy.model_management
 import numbers
+import logging
 
 RMSNorm = None
 
@@ -9,6 +10,7 @@ try:
     RMSNorm = torch.nn.RMSNorm
 except:
     rms_norm_torch = None
+    logging.warning("Please update pytorch to use native RMSNorm")
 
 
 def rms_norm(x, weight=None, eps=1e-6):

comfy/sampler_helpers.py

@@ -149,7 +149,7 @@ def cleanup_models(conds, models):
 
     cleanup_additional_models(set(control_cleanup))
 
-def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
+def prepare_model_patcher(model: ModelPatcher, conds, model_options: dict):
     '''
     Registers hooks from conds.
     '''
@@ -158,8 +158,8 @@ def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
     for k in conds:
         get_hooks_from_cond(conds[k], hooks)
     # add wrappers and callbacks from ModelPatcher to transformer_options
-    model_options["transformer_options"]["wrappers"] = comfy.patcher_extension.copy_nested_dicts(model.wrappers)
-    model_options["transformer_options"]["callbacks"] = comfy.patcher_extension.copy_nested_dicts(model.callbacks)
+    comfy.patcher_extension.merge_nested_dicts(model_options["transformer_options"].setdefault("wrappers", {}), model.wrappers, copy_dict1=False)
+    comfy.patcher_extension.merge_nested_dicts(model_options["transformer_options"].setdefault("callbacks", {}), model.callbacks, copy_dict1=False)
     # begin registering hooks
     registered = comfy.hooks.HookGroup()
     target_dict = comfy.hooks.create_target_dict(comfy.hooks.EnumWeightTarget.Model)

comfy/samplers.py

@@ -16,6 +16,8 @@ import comfy.sampler_helpers
 import comfy.model_patcher
 import comfy.patcher_extension
 import comfy.hooks
+import comfy.context_windows
+import comfy.utils
 import scipy.stats
 import numpy
 
@@ -60,7 +62,7 @@ def get_area_and_mult(conds, x_in, timestep_in):
         if "mask_strength" in conds:
             mask_strength = conds["mask_strength"]
         mask = conds['mask']
-        assert (mask.shape[1:] == x_in.shape[2:])
+        # assert (mask.shape[1:] == x_in.shape[2:])
 
         mask = mask[:input_x.shape[0]]
         if area is not None:
@@ -68,7 +70,7 @@ def get_area_and_mult(conds, x_in, timestep_in):
                 mask = mask.narrow(i + 1, area[len(dims) + i], area[i])
 
         mask = mask * mask_strength
-        mask = mask.unsqueeze(1).repeat(input_x.shape[0] // mask.shape[0], input_x.shape[1], 1, 1)
+        mask = mask.unsqueeze(1).repeat((input_x.shape[0] // mask.shape[0], input_x.shape[1]) + (1, ) * (mask.ndim - 1))
     else:
         mask = torch.ones_like(input_x)
     mult = mask * strength
@@ -198,14 +200,20 @@ def finalize_default_conds(model: 'BaseModel', hooked_to_run: dict[comfy.hooks.H
             hooked_to_run.setdefault(p.hooks, list())
             hooked_to_run[p.hooks] += [(p, i)]
 
-def calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
+def calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options: dict[str]):
+    handler: comfy.context_windows.ContextHandlerABC = model_options.get("context_handler", None)
+    if handler is None or not handler.should_use_context(model, conds, x_in, timestep, model_options):
+        return _calc_cond_batch_outer(model, conds, x_in, timestep, model_options)
+    return handler.execute(_calc_cond_batch_outer, model, conds, x_in, timestep, model_options)
+
+def _calc_cond_batch_outer(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
     executor = comfy.patcher_extension.WrapperExecutor.new_executor(
         _calc_cond_batch,
         comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.CALC_COND_BATCH, model_options, is_model_options=True)
     )
     return executor.execute(model, conds, x_in, timestep, model_options)
 
-def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
+def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
     out_conds = []
     out_counts = []
     # separate conds by matching hooks
@@ -546,7 +554,10 @@ def resolve_areas_and_cond_masks_multidim(conditions, dims, device):
             if len(mask.shape) == len(dims):
                 mask = mask.unsqueeze(0)
             if mask.shape[1:] != dims:
-                mask = torch.nn.functional.interpolate(mask.unsqueeze(1), size=dims, mode='bilinear', align_corners=False).squeeze(1)
+                if mask.ndim < 4:
+                    mask = comfy.utils.common_upscale(mask.unsqueeze(1), dims[-1], dims[-2], 'bilinear', 'none').squeeze(1)
+                else:
+                    mask = comfy.utils.common_upscale(mask, dims[-1], dims[-2], 'bilinear', 'none')
 
             if modified.get("set_area_to_bounds", False): #TODO: handle dim != 2
                 bounds = torch.max(torch.abs(mask),dim=0).values.unsqueeze(0)
@@ -718,7 +729,7 @@ class Sampler:
 
 KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2","dpm_2", "dpm_2_ancestral",
                   "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
-                  "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
+                  "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_2m_sde_heun", "dpmpp_2m_sde_heun_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
                   "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",
                   "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3", "sa_solver", "sa_solver_pece"]
 
@@ -946,7 +957,14 @@ class CFGGuider:
             self.original_conds[k] = comfy.sampler_helpers.convert_cond(conds[k])
 
     def __call__(self, *args, **kwargs):
-        return self.predict_noise(*args, **kwargs)
+        return self.outer_predict_noise(*args, **kwargs)
+
+    def outer_predict_noise(self, x, timestep, model_options={}, seed=None):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self.predict_noise,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREDICT_NOISE, self.model_options, is_model_options=True)
+        ).execute(x, timestep, model_options, seed)
 
     def predict_noise(self, x, timestep, model_options={}, seed=None):
         return sampling_function(self.inner_model, x, timestep, self.conds.get("negative", None), self.conds.get("positive", None), self.cfg, model_options=model_options, seed=seed)

comfy/sd1_clip.py

@@ -204,17 +204,19 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
             tokens_embed = self.transformer.get_input_embeddings()(tokens_embed, out_dtype=torch.float32)
             index = 0
             pad_extra = 0
+            embeds_info = []
             for o in other_embeds:
                 emb = o[1]
                 if torch.is_tensor(emb):
                     emb = {"type": "embedding", "data": emb}
 
+                extra = None
                 emb_type = emb.get("type", None)
                 if emb_type == "embedding":
                     emb = emb.get("data", None)
                 else:
                     if hasattr(self.transformer, "preprocess_embed"):
-                        emb = self.transformer.preprocess_embed(emb, device=device)
+                        emb, extra = self.transformer.preprocess_embed(emb, device=device)
                     else:
                         emb = None
 
@@ -229,6 +231,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
                     tokens_embed = torch.cat([tokens_embed[:, :ind], emb, tokens_embed[:, ind:]], dim=1)
                     attention_mask = attention_mask[:ind] + [1] * emb_shape + attention_mask[ind:]
                     index += emb_shape - 1
+                    embeds_info.append({"type": emb_type, "index": ind, "size": emb_shape, "extra": extra})
                 else:
                     index += -1
                     pad_extra += emb_shape
@@ -243,11 +246,11 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
             attention_masks.append(attention_mask)
             num_tokens.append(sum(attention_mask))
 
-        return torch.cat(embeds_out), torch.tensor(attention_masks, device=device, dtype=torch.long), num_tokens
+        return torch.cat(embeds_out), torch.tensor(attention_masks, device=device, dtype=torch.long), num_tokens, embeds_info
 
     def forward(self, tokens):
         device = self.transformer.get_input_embeddings().weight.device
-        embeds, attention_mask, num_tokens = self.process_tokens(tokens, device)
+        embeds, attention_mask, num_tokens, embeds_info = self.process_tokens(tokens, device)
 
         attention_mask_model = None
         if self.enable_attention_masks:
@@ -258,7 +261,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         else:
             intermediate_output = self.layer_idx
 
-        outputs = self.transformer(None, attention_mask_model, embeds=embeds, num_tokens=num_tokens, intermediate_output=intermediate_output, final_layer_norm_intermediate=self.layer_norm_hidden_state, dtype=torch.float32)
+        outputs = self.transformer(None, attention_mask_model, embeds=embeds, num_tokens=num_tokens, intermediate_output=intermediate_output, final_layer_norm_intermediate=self.layer_norm_hidden_state, dtype=torch.float32, embeds_info=embeds_info)
 
         if self.layer == "last":
             z = outputs[0].float()
@@ -531,7 +534,10 @@ class SDTokenizer:
         min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding)
 
         text = escape_important(text)
-        parsed_weights = token_weights(text, 1.0)
+        if kwargs.get("disable_weights", False):
+            parsed_weights = [(text, 1.0)]
+        else:
+            parsed_weights = token_weights(text, 1.0)
 
         # tokenize words
         tokens = []

comfy/supported_models.py

@@ -700,7 +700,7 @@ class Flux(supported_models_base.BASE):
     unet_extra_config = {}
     latent_format = latent_formats.Flux
 
-    memory_usage_factor = 2.8
+    memory_usage_factor = 3.1 # TODO: debug why flux mem usage is so weird on windows.
 
     supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
@@ -1046,6 +1046,18 @@ class WAN21_Camera(WAN21_T2V):
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.WAN21_Camera(self, image_to_video=False, device=device)
         return out
+
+class WAN22_Camera(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "camera_2.2",
+        "in_dim": 36,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21_Camera(self, image_to_video=False, device=device)
+        return out
+
 class WAN21_Vace(WAN21_T2V):
     unet_config = {
         "image_model": "wan2.1",
@@ -1060,6 +1072,19 @@ class WAN21_Vace(WAN21_T2V):
         out = model_base.WAN21_Vace(self, image_to_video=False, device=device)
         return out
 
+class WAN22_S2V(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "s2v",
+    }
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN22_S2V(self, device=device)
+        return out
+
 class WAN22_T2V(WAN21_T2V):
     unet_config = {
         "image_model": "wan2.1",
@@ -1260,6 +1285,6 @@ class QwenImage(supported_models_base.BASE):
         return supported_models_base.ClipTarget(comfy.text_encoders.qwen_image.QwenImageTokenizer, comfy.text_encoders.qwen_image.te(**hunyuan_detect))
 
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2, QwenImage]
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2, QwenImage]
 
 models += [SVD_img2vid]

comfy/text_encoders/bert.py

@@ -116,7 +116,7 @@ class BertModel_(torch.nn.Module):
         self.embeddings = BertEmbeddings(config_dict["vocab_size"], config_dict["max_position_embeddings"], config_dict["type_vocab_size"], config_dict["pad_token_id"], embed_dim, layer_norm_eps, dtype, device, operations)
         self.encoder = BertEncoder(config_dict["num_hidden_layers"], embed_dim, config_dict["intermediate_size"], config_dict["num_attention_heads"], layer_norm_eps, dtype, device, operations)
 
-    def forward(self, input_tokens, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
+    def forward(self, input_tokens, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]):
         x = self.embeddings(input_tokens, embeds=embeds, dtype=dtype)
         mask = None
         if attention_mask is not None:

comfy/text_encoders/llama.py

@@ -2,12 +2,14 @@ import torch
 import torch.nn as nn
 from dataclasses import dataclass
 from typing import Optional, Any
+import math
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
 import comfy.ldm.common_dit
 
 import comfy.model_management
+from . import qwen_vl
 
 @dataclass
 class Llama2Config:
@@ -25,6 +27,7 @@ class Llama2Config:
     rms_norm_add = False
     mlp_activation = "silu"
     qkv_bias = False
+    rope_dims = None
 
 @dataclass
 class Qwen25_3BConfig:
@@ -42,6 +45,7 @@ class Qwen25_3BConfig:
     rms_norm_add = False
     mlp_activation = "silu"
     qkv_bias = True
+    rope_dims = None
 
 @dataclass
 class Qwen25_7BVLI_Config:
@@ -59,6 +63,7 @@ class Qwen25_7BVLI_Config:
     rms_norm_add = False
     mlp_activation = "silu"
     qkv_bias = True
+    rope_dims = [16, 24, 24]
 
 @dataclass
 class Gemma2_2B_Config:
@@ -76,6 +81,7 @@ class Gemma2_2B_Config:
     rms_norm_add = True
     mlp_activation = "gelu_pytorch_tanh"
     qkv_bias = False
+    rope_dims = None
 
 class RMSNorm(nn.Module):
     def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None):
@@ -100,24 +106,30 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-def precompute_freqs_cis(head_dim, seq_len, theta, device=None):
+def precompute_freqs_cis(head_dim, position_ids, theta, rope_dims=None, device=None):
     theta_numerator = torch.arange(0, head_dim, 2, device=device).float()
     inv_freq = 1.0 / (theta ** (theta_numerator / head_dim))
 
-    position_ids = torch.arange(0, seq_len, device=device).unsqueeze(0)
-
     inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
     position_ids_expanded = position_ids[:, None, :].float()
     freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
     emb = torch.cat((freqs, freqs), dim=-1)
     cos = emb.cos()
     sin = emb.sin()
+    if rope_dims is not None and position_ids.shape[0] > 1:
+        mrope_section = rope_dims * 2
+        cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0)
+        sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0)
+    else:
+        cos = cos.unsqueeze(1)
+        sin = sin.unsqueeze(1)
+
     return (cos, sin)
 
 
 def apply_rope(xq, xk, freqs_cis):
-    cos = freqs_cis[0].unsqueeze(1)
-    sin = freqs_cis[1].unsqueeze(1)
+    cos = freqs_cis[0]
+    sin = freqs_cis[1]
     q_embed = (xq * cos) + (rotate_half(xq) * sin)
     k_embed = (xk * cos) + (rotate_half(xk) * sin)
     return q_embed, k_embed
@@ -277,7 +289,7 @@ class Llama2_(nn.Module):
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
         # self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype)
 
-    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
+    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[]):
         if embeds is not None:
             x = embeds
         else:
@@ -286,9 +298,13 @@ class Llama2_(nn.Module):
         if self.normalize_in:
             x *= self.config.hidden_size ** 0.5
 
+        if position_ids is None:
+            position_ids = torch.arange(0, x.shape[1], device=x.device).unsqueeze(0)
+
         freqs_cis = precompute_freqs_cis(self.config.head_dim,
-                                         x.shape[1],
+                                         position_ids,
                                          self.config.rope_theta,
+                                         self.config.rope_dims,
                                          device=x.device)
 
         mask = None
@@ -372,8 +388,38 @@ class Qwen25_7BVLI(BaseLlama, torch.nn.Module):
         self.num_layers = config.num_hidden_layers
 
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.visual = qwen_vl.Qwen2VLVisionTransformer(hidden_size=1280, output_hidden_size=config.hidden_size, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+    def preprocess_embed(self, embed, device):
+        if embed["type"] == "image":
+            image, grid = qwen_vl.process_qwen2vl_images(embed["data"])
+            return self.visual(image.to(device, dtype=torch.float32), grid), grid
+        return None, None
+
+    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]):
+        grid = None
+        for e in embeds_info:
+            if e.get("type") == "image":
+                grid = e.get("extra", None)
+                position_ids = torch.zeros((3, embeds.shape[1]), device=embeds.device)
+                start = e.get("index")
+                position_ids[:, :start] = torch.arange(0, start, device=embeds.device)
+                end = e.get("size") + start
+                len_max = int(grid.max()) // 2
+                start_next = len_max + start
+                position_ids[:, end:] = torch.arange(start_next, start_next + (embeds.shape[1] - end), device=embeds.device)
+                position_ids[0, start:end] = start
+                max_d = int(grid[0][1]) // 2
+                position_ids[1, start:end] = torch.arange(start, start + max_d, device=embeds.device).unsqueeze(1).repeat(1, math.ceil((end - start) / max_d)).flatten(0)[:end - start]
+                max_d = int(grid[0][2]) // 2
+                position_ids[2, start:end] = torch.arange(start, start + max_d, device=embeds.device).unsqueeze(0).repeat(math.ceil((end - start) / max_d), 1).flatten(0)[:end - start]
+
+        if grid is None:
+            position_ids = None
+
+        return super().forward(x, attention_mask=attention_mask, embeds=embeds, num_tokens=num_tokens, intermediate_output=intermediate_output, final_layer_norm_intermediate=final_layer_norm_intermediate, dtype=dtype, position_ids=position_ids)
+
 class Gemma2_2B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()

comfy/text_encoders/qwen_image.py

@@ -15,13 +15,27 @@ class QwenImageTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen25_7b", tokenizer=Qwen25_7BVLITokenizer)
         self.llama_template = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+        self.llama_template_images = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n"
 
-    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None,**kwargs):
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], **kwargs):
         if llama_template is None:
-            llama_text = self.llama_template.format(text)
+            if len(images) > 0:
+                llama_text = self.llama_template_images.format(text)
+            else:
+                llama_text = self.llama_template.format(text)
         else:
             llama_text = llama_template.format(text)
-        return super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, **kwargs)
+        tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
+        key_name = next(iter(tokens))
+        embed_count = 0
+        qwen_tokens = tokens[key_name]
+        for r in qwen_tokens:
+            for i in range(len(r)):
+                if r[i][0] == 151655:
+                    if len(images) > embed_count:
+                        r[i] = ({"type": "image", "data": images[embed_count], "original_type": "image"},) + r[i][1:]
+                        embed_count += 1
+        return tokens
 
 
 class Qwen25_7BVLIModel(sd1_clip.SDClipModel):

comfy/text_encoders/qwen_vl.py

@@ -0,0 +1,428 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Optional, Tuple
+import math
+from comfy.ldm.modules.attention import optimized_attention_for_device
+
+
+def process_qwen2vl_images(
+    images: torch.Tensor,
+    min_pixels: int = 3136,
+    max_pixels: int = 12845056,
+    patch_size: int = 14,
+    temporal_patch_size: int = 2,
+    merge_size: int = 2,
+    image_mean: list = None,
+    image_std: list = None,
+):
+    if image_mean is None:
+        image_mean = [0.48145466, 0.4578275, 0.40821073]
+    if image_std is None:
+        image_std = [0.26862954, 0.26130258, 0.27577711]
+
+    batch_size, height, width, channels = images.shape
+    device = images.device
+    # dtype = images.dtype
+
+    images = images.permute(0, 3, 1, 2)
+
+    grid_thw_list = []
+    img = images[0]
+
+    factor = patch_size * merge_size
+
+    h_bar = round(height / factor) * factor
+    w_bar = round(width / factor) * factor
+
+    if h_bar * w_bar > max_pixels:
+        beta = math.sqrt((height * width) / max_pixels)
+        h_bar = max(factor, math.floor(height / beta / factor) * factor)
+        w_bar = max(factor, math.floor(width / beta / factor) * factor)
+    elif h_bar * w_bar < min_pixels:
+        beta = math.sqrt(min_pixels / (height * width))
+        h_bar = math.ceil(height * beta / factor) * factor
+        w_bar = math.ceil(width * beta / factor) * factor
+
+    img_resized = F.interpolate(
+        img.unsqueeze(0),
+        size=(h_bar, w_bar),
+        mode='bilinear',
+        align_corners=False
+    ).squeeze(0)
+
+    normalized = img_resized.clone()
+    for c in range(3):
+        normalized[c] = (img_resized[c] - image_mean[c]) / image_std[c]
+
+    grid_h = h_bar // patch_size
+    grid_w = w_bar // patch_size
+    grid_thw = torch.tensor([1, grid_h, grid_w], device=device, dtype=torch.long)
+
+    pixel_values = normalized
+    grid_thw_list.append(grid_thw)
+    image_grid_thw = torch.stack(grid_thw_list)
+
+    grid_t = 1
+    channel = pixel_values.shape[0]
+    pixel_values = pixel_values.unsqueeze(0).repeat(2, 1, 1, 1)
+
+    patches = pixel_values.reshape(
+        grid_t,
+        temporal_patch_size,
+        channel,
+        grid_h // merge_size,
+        merge_size,
+        patch_size,
+        grid_w // merge_size,
+        merge_size,
+        patch_size,
+    )
+
+    patches = patches.permute(0, 3, 6, 4, 7, 2, 1, 5, 8)
+    flatten_patches = patches.reshape(
+        grid_t * grid_h * grid_w,
+        channel * temporal_patch_size * patch_size * patch_size
+    )
+
+    return flatten_patches, image_grid_thw
+
+
+class VisionPatchEmbed(nn.Module):
+    def __init__(
+        self,
+        patch_size: int = 14,
+        temporal_patch_size: int = 2,
+        in_channels: int = 3,
+        embed_dim: int = 3584,
+        device=None,
+        dtype=None,
+        ops=None,
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.temporal_patch_size = temporal_patch_size
+        self.in_channels = in_channels
+        self.embed_dim = embed_dim
+
+        kernel_size = [temporal_patch_size, patch_size, patch_size]
+        self.proj = ops.Conv3d(
+            in_channels,
+            embed_dim,
+            kernel_size=kernel_size,
+            stride=kernel_size,
+            bias=False,
+            device=device,
+            dtype=dtype
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = hidden_states.view(
+            -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
+        )
+        hidden_states = self.proj(hidden_states)
+        return hidden_states.view(-1, self.embed_dim)
+
+
+def rotate_half(x):
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb_vision(q, k, cos, sin):
+    cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float()
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class VisionRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, theta: float = 10000.0):
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+
+    def forward(self, seqlen: int, device) -> torch.Tensor:
+        inv_freq = 1.0 / (self.theta ** (torch.arange(0, self.dim, 2, dtype=torch.float, device=device) / self.dim))
+        seq = torch.arange(seqlen, device=inv_freq.device, dtype=inv_freq.dtype)
+        freqs = torch.outer(seq, inv_freq)
+        return freqs
+
+
+class PatchMerger(nn.Module):
+    def __init__(self, dim: int, context_dim: int, spatial_merge_size: int = 2, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.hidden_size = context_dim * (spatial_merge_size ** 2)
+        self.ln_q = ops.RMSNorm(context_dim, eps=1e-6, device=device, dtype=dtype)
+        self.mlp = nn.Sequential(
+            ops.Linear(self.hidden_size, self.hidden_size, device=device, dtype=dtype),
+            nn.GELU(),
+            ops.Linear(self.hidden_size, dim, device=device, dtype=dtype),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.ln_q(x).reshape(-1, self.hidden_size)
+        x = self.mlp(x)
+        return x
+
+
+class VisionAttention(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.num_heads = num_heads
+        self.head_dim = hidden_size // num_heads
+        self.scaling = self.head_dim ** -0.5
+
+        self.qkv = ops.Linear(hidden_size, hidden_size * 3, bias=True, device=device, dtype=dtype)
+        self.proj = ops.Linear(hidden_size, hidden_size, bias=True, device=device, dtype=dtype)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        cu_seqlens=None,
+        optimized_attention=None,
+    ) -> torch.Tensor:
+        if hidden_states.dim() == 2:
+            seq_length, _ = hidden_states.shape
+            batch_size = 1
+            hidden_states = hidden_states.unsqueeze(0)
+        else:
+            batch_size, seq_length, _ = hidden_states.shape
+
+        qkv = self.qkv(hidden_states)
+        qkv = qkv.reshape(batch_size, seq_length, 3, self.num_heads, self.head_dim)
+        query_states, key_states, value_states = qkv.reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
+
+        if position_embeddings is not None:
+            cos, sin = position_embeddings
+            query_states, key_states = apply_rotary_pos_emb_vision(query_states, key_states, cos, sin)
+
+        query_states = query_states.transpose(0, 1).unsqueeze(0)
+        key_states = key_states.transpose(0, 1).unsqueeze(0)
+        value_states = value_states.transpose(0, 1).unsqueeze(0)
+
+        lengths = cu_seqlens[1:] - cu_seqlens[:-1]
+        splits = [
+            torch.split(tensor, lengths.tolist(), dim=2) for tensor in (query_states, key_states, value_states)
+        ]
+
+        attn_outputs = [
+            optimized_attention(q, k, v, self.num_heads, skip_reshape=True)
+            for q, k, v in zip(*splits)
+        ]
+        attn_output = torch.cat(attn_outputs, dim=1)
+        attn_output = attn_output.reshape(seq_length, -1)
+        attn_output = self.proj(attn_output)
+
+        return attn_output
+
+
+class VisionMLP(nn.Module):
+    def __init__(self, hidden_size: int, intermediate_size: int, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.gate_proj = ops.Linear(hidden_size, intermediate_size, bias=True, device=device, dtype=dtype)
+        self.up_proj = ops.Linear(hidden_size, intermediate_size, bias=True, device=device, dtype=dtype)
+        self.down_proj = ops.Linear(intermediate_size, hidden_size, bias=True, device=device, dtype=dtype)
+        self.act_fn = nn.SiLU()
+
+    def forward(self, hidden_state):
+        return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
+
+
+class VisionBlock(nn.Module):
+    def __init__(self, hidden_size: int, intermediate_size: int, num_heads: int, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.norm1 = ops.RMSNorm(hidden_size, eps=1e-6, device=device, dtype=dtype)
+        self.norm2 = ops.RMSNorm(hidden_size, eps=1e-6, device=device, dtype=dtype)
+        self.attn = VisionAttention(hidden_size, num_heads, device=device, dtype=dtype, ops=ops)
+        self.mlp = VisionMLP(hidden_size, intermediate_size, device=device, dtype=dtype, ops=ops)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        cu_seqlens=None,
+        optimized_attention=None,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.attn(hidden_states, position_embeddings, cu_seqlens, optimized_attention)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class Qwen2VLVisionTransformer(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int = 3584,
+        output_hidden_size: int = 3584,
+        intermediate_size: int = 3420,
+        num_heads: int = 16,
+        num_layers: int = 32,
+        patch_size: int = 14,
+        temporal_patch_size: int = 2,
+        spatial_merge_size: int = 2,
+        window_size: int = 112,
+        device=None,
+        dtype=None,
+        ops=None
+    ):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.patch_size = patch_size
+        self.spatial_merge_size = spatial_merge_size
+        self.window_size = window_size
+        self.fullatt_block_indexes = [7, 15, 23, 31]
+
+        self.patch_embed = VisionPatchEmbed(
+            patch_size=patch_size,
+            temporal_patch_size=temporal_patch_size,
+            in_channels=3,
+            embed_dim=hidden_size,
+            device=device,
+            dtype=dtype,
+            ops=ops,
+        )
+
+        head_dim = hidden_size // num_heads
+        self.rotary_pos_emb = VisionRotaryEmbedding(head_dim // 2)
+
+        self.blocks = nn.ModuleList([
+            VisionBlock(hidden_size, intermediate_size, num_heads, device, dtype, ops)
+            for _ in range(num_layers)
+        ])
+
+        self.merger = PatchMerger(
+            dim=output_hidden_size,
+            context_dim=hidden_size,
+            spatial_merge_size=spatial_merge_size,
+            device=device,
+            dtype=dtype,
+            ops=ops,
+        )
+
+    def get_window_index(self, grid_thw):
+        window_index = []
+        cu_window_seqlens = [0]
+        window_index_id = 0
+        vit_merger_window_size = self.window_size // self.spatial_merge_size // self.patch_size
+
+        for grid_t, grid_h, grid_w in grid_thw:
+            llm_grid_h = grid_h // self.spatial_merge_size
+            llm_grid_w = grid_w // self.spatial_merge_size
+
+            index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(grid_t, llm_grid_h, llm_grid_w)
+
+            pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
+            pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
+            num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
+            num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
+
+            index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
+            index_padded = index_padded.reshape(
+                grid_t,
+                num_windows_h,
+                vit_merger_window_size,
+                num_windows_w,
+                vit_merger_window_size,
+            )
+            index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
+                grid_t,
+                num_windows_h * num_windows_w,
+                vit_merger_window_size,
+                vit_merger_window_size,
+            )
+
+            seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
+            index_padded = index_padded.reshape(-1)
+            index_new = index_padded[index_padded != -100]
+            window_index.append(index_new + window_index_id)
+
+            cu_seqlens_tmp = seqlens.cumsum(0) * self.spatial_merge_size * self.spatial_merge_size + cu_window_seqlens[-1]
+            cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
+            window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()
+
+        window_index = torch.cat(window_index, dim=0)
+        return window_index, cu_window_seqlens
+
+    def get_position_embeddings(self, grid_thw, device):
+        pos_ids = []
+
+        for t, h, w in grid_thw:
+            hpos_ids = torch.arange(h, device=device).unsqueeze(1).expand(-1, w)
+            hpos_ids = hpos_ids.reshape(
+                h // self.spatial_merge_size,
+                self.spatial_merge_size,
+                w // self.spatial_merge_size,
+                self.spatial_merge_size,
+            )
+            hpos_ids = hpos_ids.permute(0, 2, 1, 3).flatten()
+
+            wpos_ids = torch.arange(w, device=device).unsqueeze(0).expand(h, -1)
+            wpos_ids = wpos_ids.reshape(
+                h // self.spatial_merge_size,
+                self.spatial_merge_size,
+                w // self.spatial_merge_size,
+                self.spatial_merge_size,
+            )
+            wpos_ids = wpos_ids.permute(0, 2, 1, 3).flatten()
+
+            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+
+        pos_ids = torch.cat(pos_ids, dim=0)
+        max_grid_size = grid_thw[:, 1:].max()
+        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size, device)
+        return rotary_pos_emb_full[pos_ids].flatten(1)
+
+    def forward(
+        self,
+        pixel_values: torch.Tensor,
+        image_grid_thw: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        optimized_attention = optimized_attention_for_device(pixel_values.device, mask=False, small_input=True)
+
+        hidden_states = self.patch_embed(pixel_values)
+
+        window_index, cu_window_seqlens = self.get_window_index(image_grid_thw)
+        cu_window_seqlens = torch.tensor(cu_window_seqlens, device=hidden_states.device)
+        cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)
+
+        position_embeddings = self.get_position_embeddings(image_grid_thw, hidden_states.device)
+
+        seq_len, _ = hidden_states.size()
+        spatial_merge_unit = self.spatial_merge_size * self.spatial_merge_size
+
+        hidden_states = hidden_states.reshape(seq_len // spatial_merge_unit, spatial_merge_unit, -1)
+        hidden_states = hidden_states[window_index, :, :]
+        hidden_states = hidden_states.reshape(seq_len, -1)
+
+        position_embeddings = position_embeddings.reshape(seq_len // spatial_merge_unit, spatial_merge_unit, -1)
+        position_embeddings = position_embeddings[window_index, :, :]
+        position_embeddings = position_embeddings.reshape(seq_len, -1)
+        position_embeddings = torch.cat((position_embeddings, position_embeddings), dim=-1)
+        position_embeddings = (position_embeddings.cos(), position_embeddings.sin())
+
+        cu_seqlens = torch.repeat_interleave(image_grid_thw[:, 1] * image_grid_thw[:, 2], image_grid_thw[:, 0]).cumsum(
+            dim=0,
+            dtype=torch.int32,
+        )
+        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+
+        for i, block in enumerate(self.blocks):
+            if i in self.fullatt_block_indexes:
+                cu_seqlens_now = cu_seqlens
+            else:
+                cu_seqlens_now = cu_window_seqlens
+            hidden_states = block(hidden_states, position_embeddings, cu_seqlens_now, optimized_attention=optimized_attention)
+
+        hidden_states = self.merger(hidden_states)
+        return hidden_states

comfy/text_encoders/t5.py

@@ -199,7 +199,7 @@ class T5Stack(torch.nn.Module):
         self.final_layer_norm = T5LayerNorm(model_dim, dtype=dtype, device=device, operations=operations)
         # self.dropout = nn.Dropout(config.dropout_rate)
 
-    def forward(self, x, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
+    def forward(self, x, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]):
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])

comfy/utils.py

@@ -50,16 +50,10 @@ if hasattr(torch.serialization, "add_safe_globals"):  # TODO: this was added in
 else:
     logging.info("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended.")
 
-def is_html_file(file_path):
-    with open(file_path, "rb") as f:
-        content = f.read(100)
-        return b"<!DOCTYPE html>" in content or b"<html" in content
-
 def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
     if device is None:
         device = torch.device("cpu")
     metadata = None
-
     if ckpt.lower().endswith(".safetensors") or ckpt.lower().endswith(".sft"):
         try:
             with safetensors.safe_open(ckpt, framework="pt", device=device.type) as f:
@@ -72,8 +66,6 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
                 if return_metadata:
                     metadata = f.metadata()
         except Exception as e:
-            if is_html_file(ckpt):
-                raise ValueError("{}\n\nFile path: {}\n\nThe requested file is an HTML document not a safetensors file. Please re-download the file, not the web page.".format(e, ckpt))
             if len(e.args) > 0:
                 message = e.args[0]
                 if "HeaderTooLarge" in message:
@@ -101,13 +93,6 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
                     sd = pl_sd
             else:
                 sd = pl_sd
-
-    try:
-        from app.model_processor import model_processor
-        model_processor.process_file(ckpt)
-    except Exception as e:
-        logging.error(f"Error processing file {ckpt}: {e}")
-
     return (sd, metadata) if return_metadata else sd
 
 def save_torch_file(sd, ckpt, metadata=None):

comfy/weight_adapter/lokr.py

@@ -97,6 +97,9 @@ class LoKrAdapter(WeightAdapterBase):
             (mat1, mat2, alpha, None, None, None, None, None, None)
         )
 
+    def to_train(self):
+        return LokrDiff(self.weights)
+
     @classmethod
     def load(
         cls,

comfy_api/latest/_input_impl/video_types.py

@@ -8,6 +8,7 @@ import av
 import io
 import json
 import numpy as np
+import math
 import torch
 from comfy_api.latest._util import VideoContainer, VideoCodec, VideoComponents
 
@@ -282,8 +283,6 @@ class VideoFromComponents(VideoInput):
             if self.__components.audio:
                 audio_sample_rate = int(self.__components.audio['sample_rate'])
                 audio_stream = output.add_stream('aac', rate=audio_sample_rate)
-                audio_stream.sample_rate = audio_sample_rate
-                audio_stream.format = 'fltp'
 
             # Encode video
             for i, frame in enumerate(self.__components.images):
@@ -298,27 +297,12 @@ class VideoFromComponents(VideoInput):
             output.mux(packet)
 
             if audio_stream and self.__components.audio:
-                # Encode audio
-                samples_per_frame = int(audio_sample_rate / frame_rate)
-                num_frames = self.__components.audio['waveform'].shape[2] // samples_per_frame
-                for i in range(num_frames):
-                    start = i * samples_per_frame
-                    end = start + samples_per_frame
-                    # TODO(Feature) - Add support for stereo audio
-                    chunk = (
-                        self.__components.audio["waveform"][0, 0, start:end]
-                        .unsqueeze(0)
-                        .contiguous()
-                        .numpy()
-                    )
-                    audio_frame = av.AudioFrame.from_ndarray(chunk, format='fltp', layout='mono')
-                    audio_frame.sample_rate = audio_sample_rate
-                    audio_frame.pts = i * samples_per_frame
-                    for packet in audio_stream.encode(audio_frame):
-                        output.mux(packet)
-
-                # Flush audio
-                for packet in audio_stream.encode(None):
-                    output.mux(packet)
-
+                waveform = self.__components.audio['waveform']
+                waveform = waveform[:, :, :math.ceil((audio_sample_rate / frame_rate) * self.__components.images.shape[0])]
+                frame = av.AudioFrame.from_ndarray(waveform.movedim(2, 1).reshape(1, -1).float().numpy(), format='flt', layout='mono' if waveform.shape[1] == 1 else 'stereo')
+                frame.sample_rate = audio_sample_rate
+                frame.pts = 0
+                output.mux(audio_stream.encode(frame))
 
+                # Flush encoder
+                output.mux(audio_stream.encode(None))

comfy_api/latest/_io.py

@@ -726,6 +726,18 @@ class SEGS(ComfyTypeIO):
 class AnyType(ComfyTypeIO):
     Type = Any
 
+@comfytype(io_type="MODEL_PATCH")
+class MODEL_PATCH(ComfyTypeIO):
+    Type = Any
+
+@comfytype(io_type="AUDIO_ENCODER")
+class AudioEncoder(ComfyTypeIO):
+    Type = Any
+
+@comfytype(io_type="AUDIO_ENCODER_OUTPUT")
+class AudioEncoderOutput(ComfyTypeIO):
+    Type = Any
+
 @comfytype(io_type="COMFY_MULTITYPED_V3")
 class MultiType:
     Type = Any
@@ -1580,6 +1592,7 @@ class _IO:
     Model = Model
     ClipVision = ClipVision
     ClipVisionOutput = ClipVisionOutput
+    AudioEncoderOutput = AudioEncoderOutput
     StyleModel = StyleModel
     Gligen = Gligen
     UpscaleModel = UpscaleModel

comfy_api/latest/_ui.py

@@ -12,7 +12,7 @@ import torch
 try:
     import torchaudio
     TORCH_AUDIO_AVAILABLE = True
-except ImportError:
+except:
     TORCH_AUDIO_AVAILABLE = False
 from PIL import Image as PILImage
 from PIL.PngImagePlugin import PngInfo

comfy_api_nodes/apis/__init__.py

@@ -1315,6 +1315,7 @@ class KlingTaskStatus(str, Enum):
 class KlingTextToVideoModelName(str, Enum):
     kling_v1 = 'kling-v1'
     kling_v1_6 = 'kling-v1-6'
+    kling_v2_1_master = 'kling-v2-1-master'
 
 
 class KlingVideoGenAspectRatio(str, Enum):
@@ -1347,6 +1348,8 @@ class KlingVideoGenModelName(str, Enum):
     kling_v1_5 = 'kling-v1-5'
     kling_v1_6 = 'kling-v1-6'
     kling_v2_master = 'kling-v2-master'
+    kling_v2_1 = 'kling-v2-1'
+    kling_v2_1_master = 'kling-v2-1-master'
 
 
 class KlingVideoResult(BaseModel):
@@ -1620,13 +1623,14 @@ class MinimaxTaskResultResponse(BaseModel):
     task_id: str = Field(..., description='The task ID being queried.')
 
 
-class Model(str, Enum):
+class MiniMaxModel(str, Enum):
     T2V_01_Director = 'T2V-01-Director'
     I2V_01_Director = 'I2V-01-Director'
     S2V_01 = 'S2V-01'
     I2V_01 = 'I2V-01'
     I2V_01_live = 'I2V-01-live'
     T2V_01 = 'T2V-01'
+    Hailuo_02 = 'MiniMax-Hailuo-02'
 
 
 class SubjectReferenceItem(BaseModel):
@@ -1648,7 +1652,7 @@ class MinimaxVideoGenerationRequest(BaseModel):
         None,
         description='URL or base64 encoding of the first frame image. Required when model is I2V-01, I2V-01-Director, or I2V-01-live.',
     )
-    model: Model = Field(
+    model: MiniMaxModel = Field(
         ...,
         description='Required. ID of model. Options: T2V-01-Director, I2V-01-Director, S2V-01, I2V-01, I2V-01-live, T2V-01',
     )
@@ -1665,6 +1669,14 @@ class MinimaxVideoGenerationRequest(BaseModel):
         None,
         description='Only available when model is S2V-01. The model will generate a video based on the subject uploaded through this parameter.',
     )
+    duration: Optional[int] = Field(
+        None,
+        description="The length of the output video in seconds."
+    )
+    resolution: Optional[str] = Field(
+        None,
+        description="The dimensions of the video display. 1080p corresponds to 1920 x 1080 pixels, 768p corresponds to 1366 x 768 pixels."
+    )
 
 
 class MinimaxVideoGenerationResponse(BaseModel):

comfy_api_nodes/apis/gemini_api.py

@@ -0,0 +1,19 @@
+from __future__ import annotations
+
+from typing import List, Optional
+
+from comfy_api_nodes.apis import GeminiGenerationConfig, GeminiContent, GeminiSafetySetting, GeminiSystemInstructionContent, GeminiTool, GeminiVideoMetadata
+from pydantic import BaseModel
+
+
+class GeminiImageGenerationConfig(GeminiGenerationConfig):
+    responseModalities: Optional[List[str]] = None
+
+
+class GeminiImageGenerateContentRequest(BaseModel):
+    contents: List[GeminiContent]
+    generationConfig: Optional[GeminiImageGenerationConfig] = None
+    safetySettings: Optional[List[GeminiSafetySetting]] = None
+    systemInstruction: Optional[GeminiSystemInstructionContent] = None
+    tools: Optional[List[GeminiTool]] = None
+    videoMetadata: Optional[GeminiVideoMetadata] = None

comfy_api_nodes/nodes_gemini.py

@@ -4,8 +4,12 @@ See: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/infer
 """
 from __future__ import annotations
 
-
+import json
+import time
 import os
+import uuid
+import base64
+from io import BytesIO
 from enum import Enum
 from typing import Optional, Literal
 
@@ -22,6 +26,7 @@ from comfy_api_nodes.apis import (
     GeminiPart,
     GeminiMimeType,
 )
+from comfy_api_nodes.apis.gemini_api import GeminiImageGenerationConfig, GeminiImageGenerateContentRequest
 from comfy_api_nodes.apis.client import (
     ApiEndpoint,
     HttpMethod,
@@ -32,6 +37,7 @@ from comfy_api_nodes.apinode_utils import (
     audio_to_base64_string,
     video_to_base64_string,
     tensor_to_base64_string,
+    bytesio_to_image_tensor,
 )
 
 
@@ -46,6 +52,16 @@ class GeminiModel(str, Enum):
 
     gemini_2_5_pro_preview_05_06 = "gemini-2.5-pro-preview-05-06"
     gemini_2_5_flash_preview_04_17 = "gemini-2.5-flash-preview-04-17"
+    gemini_2_5_pro = "gemini-2.5-pro"
+    gemini_2_5_flash = "gemini-2.5-flash"
+
+
+class GeminiImageModel(str, Enum):
+    """
+    Gemini Image Model Names allowed by comfy-api
+    """
+
+    gemini_2_5_flash_image_preview = "gemini-2.5-flash-image-preview"
 
 
 def get_gemini_endpoint(
@@ -70,6 +86,135 @@ def get_gemini_endpoint(
     )
 
 
+def get_gemini_image_endpoint(
+    model: GeminiImageModel,
+) -> ApiEndpoint[GeminiGenerateContentRequest, GeminiGenerateContentResponse]:
+    """
+    Get the API endpoint for a given Gemini model.
+
+    Args:
+        model: The Gemini model to use, either as enum or string value.
+
+    Returns:
+        ApiEndpoint configured for the specific Gemini model.
+    """
+    if isinstance(model, str):
+        model = GeminiImageModel(model)
+    return ApiEndpoint(
+        path=f"{GEMINI_BASE_ENDPOINT}/{model.value}",
+        method=HttpMethod.POST,
+        request_model=GeminiImageGenerateContentRequest,
+        response_model=GeminiGenerateContentResponse,
+    )
+
+
+def create_image_parts(image_input: torch.Tensor) -> list[GeminiPart]:
+    """
+    Convert image tensor input to Gemini API compatible parts.
+
+    Args:
+        image_input: Batch of image tensors from ComfyUI.
+
+    Returns:
+        List of GeminiPart objects containing the encoded images.
+    """
+    image_parts: list[GeminiPart] = []
+    for image_index in range(image_input.shape[0]):
+        image_as_b64 = tensor_to_base64_string(
+            image_input[image_index].unsqueeze(0)
+        )
+        image_parts.append(
+            GeminiPart(
+                inlineData=GeminiInlineData(
+                    mimeType=GeminiMimeType.image_png,
+                    data=image_as_b64,
+                )
+            )
+        )
+    return image_parts
+
+
+def create_text_part(text: str) -> GeminiPart:
+    """
+    Create a text part for the Gemini API request.
+
+    Args:
+        text: The text content to include in the request.
+
+    Returns:
+        A GeminiPart object with the text content.
+    """
+    return GeminiPart(text=text)
+
+
+def get_parts_from_response(
+    response: GeminiGenerateContentResponse
+) -> list[GeminiPart]:
+    """
+    Extract all parts from the Gemini API response.
+
+    Args:
+        response: The API response from Gemini.
+
+    Returns:
+        List of response parts from the first candidate.
+    """
+    return response.candidates[0].content.parts
+
+
+def get_parts_by_type(
+    response: GeminiGenerateContentResponse, part_type: Literal["text"] | str
+) -> list[GeminiPart]:
+    """
+    Filter response parts by their type.
+
+    Args:
+        response: The API response from Gemini.
+        part_type: Type of parts to extract ("text" or a MIME type).
+
+    Returns:
+        List of response parts matching the requested type.
+    """
+    parts = []
+    for part in get_parts_from_response(response):
+        if part_type == "text" and hasattr(part, "text") and part.text:
+            parts.append(part)
+        elif (
+            hasattr(part, "inlineData")
+            and part.inlineData
+            and part.inlineData.mimeType == part_type
+        ):
+            parts.append(part)
+        # Skip parts that don't match the requested type
+    return parts
+
+
+def get_text_from_response(response: GeminiGenerateContentResponse) -> str:
+    """
+    Extract and concatenate all text parts from the response.
+
+    Args:
+        response: The API response from Gemini.
+
+    Returns:
+        Combined text from all text parts in the response.
+    """
+    parts = get_parts_by_type(response, "text")
+    return "\n".join([part.text for part in parts])
+
+
+def get_image_from_response(response: GeminiGenerateContentResponse) -> torch.Tensor:
+    image_tensors: list[torch.Tensor] = []
+    parts = get_parts_by_type(response, "image/png")
+    for part in parts:
+        image_data = base64.b64decode(part.inlineData.data)
+        returned_image = bytesio_to_image_tensor(BytesIO(image_data))
+        image_tensors.append(returned_image)
+    if len(image_tensors) == 0:
+        return torch.zeros((1,1024,1024,4))
+    return torch.cat(image_tensors, dim=0)
+
+
 class GeminiNode(ComfyNodeABC):
     """
     Node to generate text responses from a Gemini model.
@@ -97,7 +242,7 @@ class GeminiNode(ComfyNodeABC):
                     {
                         "tooltip": "The Gemini model to use for generating responses.",
                         "options": [model.value for model in GeminiModel],
-                        "default": GeminiModel.gemini_2_5_pro_preview_05_06.value,
+                        "default": GeminiModel.gemini_2_5_pro.value,
                     },
                 ),
                 "seed": (
@@ -154,59 +299,6 @@ class GeminiNode(ComfyNodeABC):
     CATEGORY = "api node/text/Gemini"
     API_NODE = True
 
-    def get_parts_from_response(
-        self, response: GeminiGenerateContentResponse
-    ) -> list[GeminiPart]:
-        """
-        Extract all parts from the Gemini API response.
-
-        Args:
-            response: The API response from Gemini.
-
-        Returns:
-            List of response parts from the first candidate.
-        """
-        return response.candidates[0].content.parts
-
-    def get_parts_by_type(
-        self, response: GeminiGenerateContentResponse, part_type: Literal["text"] | str
-    ) -> list[GeminiPart]:
-        """
-        Filter response parts by their type.
-
-        Args:
-            response: The API response from Gemini.
-            part_type: Type of parts to extract ("text" or a MIME type).
-
-        Returns:
-            List of response parts matching the requested type.
-        """
-        parts = []
-        for part in self.get_parts_from_response(response):
-            if part_type == "text" and hasattr(part, "text") and part.text:
-                parts.append(part)
-            elif (
-                hasattr(part, "inlineData")
-                and part.inlineData
-                and part.inlineData.mimeType == part_type
-            ):
-                parts.append(part)
-            # Skip parts that don't match the requested type
-        return parts
-
-    def get_text_from_response(self, response: GeminiGenerateContentResponse) -> str:
-        """
-        Extract and concatenate all text parts from the response.
-
-        Args:
-            response: The API response from Gemini.
-
-        Returns:
-            Combined text from all text parts in the response.
-        """
-        parts = self.get_parts_by_type(response, "text")
-        return "\n".join([part.text for part in parts])
-
     def create_video_parts(self, video_input: IO.VIDEO, **kwargs) -> list[GeminiPart]:
         """
         Convert video input to Gemini API compatible parts.
@@ -266,43 +358,6 @@ class GeminiNode(ComfyNodeABC):
             )
         return audio_parts
 
-    def create_image_parts(self, image_input: torch.Tensor) -> list[GeminiPart]:
-        """
-        Convert image tensor input to Gemini API compatible parts.
-
-        Args:
-            image_input: Batch of image tensors from ComfyUI.
-
-        Returns:
-            List of GeminiPart objects containing the encoded images.
-        """
-        image_parts: list[GeminiPart] = []
-        for image_index in range(image_input.shape[0]):
-            image_as_b64 = tensor_to_base64_string(
-                image_input[image_index].unsqueeze(0)
-            )
-            image_parts.append(
-                GeminiPart(
-                    inlineData=GeminiInlineData(
-                        mimeType=GeminiMimeType.image_png,
-                        data=image_as_b64,
-                    )
-                )
-            )
-        return image_parts
-
-    def create_text_part(self, text: str) -> GeminiPart:
-        """
-        Create a text part for the Gemini API request.
-
-        Args:
-            text: The text content to include in the request.
-
-        Returns:
-            A GeminiPart object with the text content.
-        """
-        return GeminiPart(text=text)
-
     async def api_call(
         self,
         prompt: str,
@@ -318,11 +373,11 @@ class GeminiNode(ComfyNodeABC):
         validate_string(prompt, strip_whitespace=False)
 
         # Create parts list with text prompt as the first part
-        parts: list[GeminiPart] = [self.create_text_part(prompt)]
+        parts: list[GeminiPart] = [create_text_part(prompt)]
 
         # Add other modal parts
         if images is not None:
-            image_parts = self.create_image_parts(images)
+            image_parts = create_image_parts(images)
             parts.extend(image_parts)
         if audio is not None:
             parts.extend(self.create_audio_parts(audio))
@@ -346,9 +401,29 @@ class GeminiNode(ComfyNodeABC):
         ).execute()
 
         # Get result output
-        output_text = self.get_text_from_response(response)
+        output_text = get_text_from_response(response)
         if unique_id and output_text:
-            PromptServer.instance.send_progress_text(output_text, node_id=unique_id)
+            # Not a true chat history like the OpenAI Chat node. It is emulated so the frontend can show a copy button.
+            render_spec = {
+                "node_id": unique_id,
+                "component": "ChatHistoryWidget",
+                "props": {
+                    "history": json.dumps(
+                        [
+                            {
+                                "prompt": prompt,
+                                "response": output_text,
+                                "response_id": str(uuid.uuid4()),
+                                "timestamp": time.time(),
+                            }
+                        ]
+                    ),
+                },
+            }
+            PromptServer.instance.send_sync(
+                "display_component",
+                render_spec,
+            )
 
         return (output_text or "Empty response from Gemini model...",)
 
@@ -437,12 +512,162 @@ class GeminiInputFiles(ComfyNodeABC):
         return (files,)
 
 
+class GeminiImage(ComfyNodeABC):
+    """
+    Node to generate text and image responses from a Gemini model.
+
+    This node allows users to interact with Google's Gemini AI models, providing
+    multimodal inputs (text, images, files) to generate coherent
+    text and image responses. The node works with the latest Gemini models, handling the
+    API communication and response parsing.
+    """
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {
+                "prompt": (
+                    IO.STRING,
+                    {
+                        "multiline": True,
+                        "default": "",
+                        "tooltip": "Text prompt for generation",
+                    },
+                ),
+                "model": (
+                    IO.COMBO,
+                    {
+                        "tooltip": "The Gemini model to use for generating responses.",
+                        "options": [model.value for model in GeminiImageModel],
+                        "default": GeminiImageModel.gemini_2_5_flash_image_preview.value,
+                    },
+                ),
+                "seed": (
+                    IO.INT,
+                    {
+                        "default": 42,
+                        "min": 0,
+                        "max": 0xFFFFFFFFFFFFFFFF,
+                        "control_after_generate": True,
+                        "tooltip": "When seed is fixed to a specific value, the model makes a best effort to provide the same response for repeated requests. Deterministic output isn't guaranteed. Also, changing the model or parameter settings, such as the temperature, can cause variations in the response even when you use the same seed value. By default, a random seed value is used.",
+                    },
+                ),
+            },
+            "optional": {
+                "images": (
+                    IO.IMAGE,
+                    {
+                        "default": None,
+                        "tooltip": "Optional image(s) to use as context for the model. To include multiple images, you can use the Batch Images node.",
+                    },
+                ),
+                "files": (
+                    "GEMINI_INPUT_FILES",
+                    {
+                        "default": None,
+                        "tooltip": "Optional file(s) to use as context for the model. Accepts inputs from the Gemini Generate Content Input Files node.",
+                    },
+                ),
+                # TODO: later we can add this parameter later
+                # "n": (
+                #     IO.INT,
+                #     {
+                #         "default": 1,
+                #         "min": 1,
+                #         "max": 8,
+                #         "step": 1,
+                #         "display": "number",
+                #         "tooltip": "How many images to generate",
+                #     },
+                # ),
+            },
+            "hidden": {
+                "auth_token": "AUTH_TOKEN_COMFY_ORG",
+                "comfy_api_key": "API_KEY_COMFY_ORG",
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+
+    RETURN_TYPES = (IO.IMAGE, IO.STRING)
+    FUNCTION = "api_call"
+    CATEGORY = "api node/image/Gemini"
+    DESCRIPTION = "Edit images synchronously via Google API."
+    API_NODE = True
+
+    async def api_call(
+        self,
+        prompt: str,
+        model: GeminiImageModel,
+        images: Optional[IO.IMAGE] = None,
+        files: Optional[list[GeminiPart]] = None,
+        n=1,
+        unique_id: Optional[str] = None,
+        **kwargs,
+    ):
+        # Validate inputs
+        validate_string(prompt, strip_whitespace=True, min_length=1)
+        # Create parts list with text prompt as the first part
+        parts: list[GeminiPart] = [create_text_part(prompt)]
+
+        # Add other modal parts
+        if images is not None:
+            image_parts = create_image_parts(images)
+            parts.extend(image_parts)
+        if files is not None:
+            parts.extend(files)
+
+        response = await SynchronousOperation(
+            endpoint=get_gemini_image_endpoint(model),
+            request=GeminiImageGenerateContentRequest(
+                contents=[
+                    GeminiContent(
+                        role="user",
+                        parts=parts,
+                    ),
+                ],
+                generationConfig=GeminiImageGenerationConfig(
+                    responseModalities=["TEXT","IMAGE"]
+                )
+            ),
+            auth_kwargs=kwargs,
+        ).execute()
+
+        output_image = get_image_from_response(response)
+        output_text = get_text_from_response(response)
+        if unique_id and output_text:
+            # Not a true chat history like the OpenAI Chat node. It is emulated so the frontend can show a copy button.
+            render_spec = {
+                "node_id": unique_id,
+                "component": "ChatHistoryWidget",
+                "props": {
+                    "history": json.dumps(
+                        [
+                            {
+                                "prompt": prompt,
+                                "response": output_text,
+                                "response_id": str(uuid.uuid4()),
+                                "timestamp": time.time(),
+                            }
+                        ]
+                    ),
+                },
+            }
+            PromptServer.instance.send_sync(
+                "display_component",
+                render_spec,
+            )
+
+        output_text = output_text or "Empty response from Gemini model..."
+        return (output_image, output_text,)
+
+
 NODE_CLASS_MAPPINGS = {
     "GeminiNode": GeminiNode,
+    "GeminiImageNode": GeminiImage,
     "GeminiInputFiles": GeminiInputFiles,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
     "GeminiNode": "Google Gemini",
+    "GeminiImageNode": "Google Gemini Image",
     "GeminiInputFiles": "Gemini Input Files",
 }

comfy_api_nodes/nodes_ideogram.py

@@ -1,8 +1,8 @@
-from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeDict
-from inspect import cleandoc
+from io import BytesIO
+from typing_extensions import override
+from comfy_api.latest import ComfyExtension, io as comfy_io
 from PIL import Image
 import numpy as np
-import io
 import torch
 from comfy_api_nodes.apis import (
     IdeogramGenerateRequest,
@@ -246,90 +246,81 @@ def display_image_urls_on_node(image_urls, node_id):
             PromptServer.instance.send_progress_text(urls_text, node_id)
 
 
-class IdeogramV1(ComfyNodeABC):
-    """
-    Generates images using the Ideogram V1 model.
-    """
-
-    def __init__(self):
-        pass
+class IdeogramV1(comfy_io.ComfyNode):
 
     @classmethod
-    def INPUT_TYPES(cls) -> InputTypeDict:
-        return {
-            "required": {
-                "prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Prompt for the image generation",
-                    },
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="IdeogramV1",
+            display_name="Ideogram V1",
+            category="api node/image/Ideogram",
+            description="Generates images using the Ideogram V1 model.",
+            inputs=[
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Prompt for the image generation",
                 ),
-                "turbo": (
-                    IO.BOOLEAN,
-                    {
-                        "default": False,
-                        "tooltip": "Whether to use turbo mode (faster generation, potentially lower quality)",
-                    }
+                comfy_io.Boolean.Input(
+                    "turbo",
+                    default=False,
+                    tooltip="Whether to use turbo mode (faster generation, potentially lower quality)",
                 ),
-            },
-            "optional": {
-                "aspect_ratio": (
-                    IO.COMBO,
-                    {
-                        "options": list(V1_V2_RATIO_MAP.keys()),
-                        "default": "1:1",
-                        "tooltip": "The aspect ratio for image generation.",
-                    },
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=list(V1_V2_RATIO_MAP.keys()),
+                    default="1:1",
+                    tooltip="The aspect ratio for image generation.",
+                    optional=True,
                 ),
-                "magic_prompt_option": (
-                    IO.COMBO,
-                    {
-                        "options": ["AUTO", "ON", "OFF"],
-                        "default": "AUTO",
-                        "tooltip": "Determine if MagicPrompt should be used in generation",
-                    },
+                comfy_io.Combo.Input(
+                    "magic_prompt_option",
+                    options=["AUTO", "ON", "OFF"],
+                    default="AUTO",
+                    tooltip="Determine if MagicPrompt should be used in generation",
+                    optional=True,
                 ),
-                "seed": (
-                    IO.INT,
-                    {
-                        "default": 0,
-                        "min": 0,
-                        "max": 2147483647,
-                        "step": 1,
-                        "control_after_generate": True,
-                        "display": "number",
-                    },
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    control_after_generate=True,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    optional=True,
                 ),
-                "negative_prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Description of what to exclude from the image",
-                    },
+                comfy_io.String.Input(
+                    "negative_prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Description of what to exclude from the image",
+                    optional=True,
                 ),
-                "num_images": (
-                    IO.INT,
-                    {"default": 1, "min": 1, "max": 8, "step": 1, "display": "number"},
+                comfy_io.Int.Input(
+                    "num_images",
+                    default=1,
+                    min=1,
+                    max=8,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    optional=True,
                 ),
-            },
-            "hidden": {
-                "auth_token": "AUTH_TOKEN_COMFY_ORG",
-                "comfy_api_key": "API_KEY_COMFY_ORG",
-                "unique_id": "UNIQUE_ID",
-            },
-        }
+            ],
+            outputs=[
+                comfy_io.Image.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+        )
 
-    RETURN_TYPES = (IO.IMAGE,)
-    FUNCTION = "api_call"
-    CATEGORY = "api node/image/Ideogram"
-    DESCRIPTION = cleandoc(__doc__ or "")
-    API_NODE = True
-
-    async def api_call(
-        self,
+    @classmethod
+    async def execute(
+        cls,
         prompt,
         turbo=False,
         aspect_ratio="1:1",
@@ -337,13 +328,15 @@ class IdeogramV1(ComfyNodeABC):
         seed=0,
         negative_prompt="",
         num_images=1,
-        unique_id=None,
-        **kwargs,
     ):
         # Determine the model based on turbo setting
         aspect_ratio = V1_V2_RATIO_MAP.get(aspect_ratio, None)
         model = "V_1_TURBO" if turbo else "V_1"
 
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
         operation = SynchronousOperation(
             endpoint=ApiEndpoint(
                 path="/proxy/ideogram/generate",
@@ -364,7 +357,7 @@ class IdeogramV1(ComfyNodeABC):
                     negative_prompt=negative_prompt if negative_prompt else None,
                 )
             ),
-            auth_kwargs=kwargs,
+            auth_kwargs=auth,
         )
 
         response = await operation.execute()
@@ -377,93 +370,85 @@ class IdeogramV1(ComfyNodeABC):
         if not image_urls:
             raise Exception("No image URLs were generated in the response")
 
-        display_image_urls_on_node(image_urls, unique_id)
-        return (await download_and_process_images(image_urls),)
+        display_image_urls_on_node(image_urls, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_and_process_images(image_urls))
 
 
-class IdeogramV2(ComfyNodeABC):
-    """
-    Generates images using the Ideogram V2 model.
-    """
-
-    def __init__(self):
-        pass
+class IdeogramV2(comfy_io.ComfyNode):
 
     @classmethod
-    def INPUT_TYPES(cls) -> InputTypeDict:
-        return {
-            "required": {
-                "prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Prompt for the image generation",
-                    },
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="IdeogramV2",
+            display_name="Ideogram V2",
+            category="api node/image/Ideogram",
+            description="Generates images using the Ideogram V2 model.",
+            inputs=[
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Prompt for the image generation",
                 ),
-                "turbo": (
-                    IO.BOOLEAN,
-                    {
-                        "default": False,
-                        "tooltip": "Whether to use turbo mode (faster generation, potentially lower quality)",
-                    }
+                comfy_io.Boolean.Input(
+                    "turbo",
+                    default=False,
+                    tooltip="Whether to use turbo mode (faster generation, potentially lower quality)",
                 ),
-            },
-            "optional": {
-                "aspect_ratio": (
-                    IO.COMBO,
-                    {
-                        "options": list(V1_V2_RATIO_MAP.keys()),
-                        "default": "1:1",
-                        "tooltip": "The aspect ratio for image generation. Ignored if resolution is not set to AUTO.",
-                    },
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=list(V1_V2_RATIO_MAP.keys()),
+                    default="1:1",
+                    tooltip="The aspect ratio for image generation. Ignored if resolution is not set to AUTO.",
+                    optional=True,
                 ),
-                "resolution": (
-                    IO.COMBO,
-                    {
-                        "options": list(V1_V1_RES_MAP.keys()),
-                        "default": "Auto",
-                        "tooltip": "The resolution for image generation. If not set to AUTO, this overrides the aspect_ratio setting.",
-                    },
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=list(V1_V1_RES_MAP.keys()),
+                    default="Auto",
+                    tooltip="The resolution for image generation. "
+                            "If not set to AUTO, this overrides the aspect_ratio setting.",
+                    optional=True,
                 ),
-                "magic_prompt_option": (
-                    IO.COMBO,
-                    {
-                        "options": ["AUTO", "ON", "OFF"],
-                        "default": "AUTO",
-                        "tooltip": "Determine if MagicPrompt should be used in generation",
-                    },
+                comfy_io.Combo.Input(
+                    "magic_prompt_option",
+                    options=["AUTO", "ON", "OFF"],
+                    default="AUTO",
+                    tooltip="Determine if MagicPrompt should be used in generation",
+                    optional=True,
                 ),
-                "seed": (
-                    IO.INT,
-                    {
-                        "default": 0,
-                        "min": 0,
-                        "max": 2147483647,
-                        "step": 1,
-                        "control_after_generate": True,
-                        "display": "number",
-                    },
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    control_after_generate=True,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    optional=True,
                 ),
-                "style_type": (
-                    IO.COMBO,
-                    {
-                        "options": ["AUTO", "GENERAL", "REALISTIC", "DESIGN", "RENDER_3D", "ANIME"],
-                        "default": "NONE",
-                        "tooltip": "Style type for generation (V2 only)",
-                    },
+                comfy_io.Combo.Input(
+                    "style_type",
+                    options=["AUTO", "GENERAL", "REALISTIC", "DESIGN", "RENDER_3D", "ANIME"],
+                    default="NONE",
+                    tooltip="Style type for generation (V2 only)",
+                    optional=True,
                 ),
-                "negative_prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Description of what to exclude from the image",
-                    },
+                comfy_io.String.Input(
+                    "negative_prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Description of what to exclude from the image",
+                    optional=True,
                 ),
-                "num_images": (
-                    IO.INT,
-                    {"default": 1, "min": 1, "max": 8, "step": 1, "display": "number"},
+                comfy_io.Int.Input(
+                    "num_images",
+                    default=1,
+                    min=1,
+                    max=8,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    optional=True,
                 ),
                 #"color_palette": (
                 #    IO.STRING,
@@ -473,22 +458,20 @@ class IdeogramV2(ComfyNodeABC):
                 #        "tooltip": "Color palette preset name or hex colors with weights",
                 #    },
                 #),
-            },
-            "hidden": {
-                "auth_token": "AUTH_TOKEN_COMFY_ORG",
-                "comfy_api_key": "API_KEY_COMFY_ORG",
-                "unique_id": "UNIQUE_ID",
-            },
-        }
+            ],
+            outputs=[
+                comfy_io.Image.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+        )
 
-    RETURN_TYPES = (IO.IMAGE,)
-    FUNCTION = "api_call"
-    CATEGORY = "api node/image/Ideogram"
-    DESCRIPTION = cleandoc(__doc__ or "")
-    API_NODE = True
-
-    async def api_call(
-        self,
+    @classmethod
+    async def execute(
+        cls,
         prompt,
         turbo=False,
         aspect_ratio="1:1",
@@ -499,8 +482,6 @@ class IdeogramV2(ComfyNodeABC):
         negative_prompt="",
         num_images=1,
         color_palette="",
-        unique_id=None,
-        **kwargs,
     ):
         aspect_ratio = V1_V2_RATIO_MAP.get(aspect_ratio, None)
         resolution = V1_V1_RES_MAP.get(resolution, None)
@@ -517,6 +498,10 @@ class IdeogramV2(ComfyNodeABC):
         else:
             final_aspect_ratio = aspect_ratio if aspect_ratio != "ASPECT_1_1" else None
 
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
         operation = SynchronousOperation(
             endpoint=ApiEndpoint(
                 path="/proxy/ideogram/generate",
@@ -540,7 +525,7 @@ class IdeogramV2(ComfyNodeABC):
                     color_palette=color_palette if color_palette else None,
                 )
             ),
-            auth_kwargs=kwargs,
+            auth_kwargs=auth,
         )
 
         response = await operation.execute()
@@ -553,108 +538,99 @@ class IdeogramV2(ComfyNodeABC):
         if not image_urls:
             raise Exception("No image URLs were generated in the response")
 
-        display_image_urls_on_node(image_urls, unique_id)
-        return (await download_and_process_images(image_urls),)
+        display_image_urls_on_node(image_urls, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_and_process_images(image_urls))
 
-class IdeogramV3(ComfyNodeABC):
-    """
-    Generates images using the Ideogram V3 model. Supports both regular image generation from text prompts and image editing with mask.
-    """
 
-    def __init__(self):
-        pass
+class IdeogramV3(comfy_io.ComfyNode):
 
     @classmethod
-    def INPUT_TYPES(cls) -> InputTypeDict:
-        return {
-            "required": {
-                "prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Prompt for the image generation or editing",
-                    },
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="IdeogramV3",
+            display_name="Ideogram V3",
+            category="api node/image/Ideogram",
+            description="Generates images using the Ideogram V3 model. "
+                        "Supports both regular image generation from text prompts and image editing with mask.",
+            inputs=[
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Prompt for the image generation or editing",
                 ),
-            },
-            "optional": {
-                "image": (
-                    IO.IMAGE,
-                    {
-                        "default": None,
-                        "tooltip": "Optional reference image for image editing.",
-                    },
+                comfy_io.Image.Input(
+                    "image",
+                    tooltip="Optional reference image for image editing.",
+                    optional=True,
                 ),
-                "mask": (
-                    IO.MASK,
-                    {
-                        "default": None,
-                        "tooltip": "Optional mask for inpainting (white areas will be replaced)",
-                    },
+                comfy_io.Mask.Input(
+                    "mask",
+                    tooltip="Optional mask for inpainting (white areas will be replaced)",
+                    optional=True,
                 ),
-                "aspect_ratio": (
-                    IO.COMBO,
-                    {
-                        "options": list(V3_RATIO_MAP.keys()),
-                        "default": "1:1",
-                        "tooltip": "The aspect ratio for image generation. Ignored if resolution is not set to Auto.",
-                    },
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=list(V3_RATIO_MAP.keys()),
+                    default="1:1",
+                    tooltip="The aspect ratio for image generation. Ignored if resolution is not set to Auto.",
+                    optional=True,
                 ),
-                "resolution": (
-                    IO.COMBO,
-                    {
-                        "options": V3_RESOLUTIONS,
-                        "default": "Auto",
-                        "tooltip": "The resolution for image generation. If not set to Auto, this overrides the aspect_ratio setting.",
-                    },
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=V3_RESOLUTIONS,
+                    default="Auto",
+                    tooltip="The resolution for image generation. "
+                            "If not set to Auto, this overrides the aspect_ratio setting.",
+                    optional=True,
                 ),
-                "magic_prompt_option": (
-                    IO.COMBO,
-                    {
-                        "options": ["AUTO", "ON", "OFF"],
-                        "default": "AUTO",
-                        "tooltip": "Determine if MagicPrompt should be used in generation",
-                    },
+                comfy_io.Combo.Input(
+                    "magic_prompt_option",
+                    options=["AUTO", "ON", "OFF"],
+                    default="AUTO",
+                    tooltip="Determine if MagicPrompt should be used in generation",
+                    optional=True,
                 ),
-                "seed": (
-                    IO.INT,
-                    {
-                        "default": 0,
-                        "min": 0,
-                        "max": 2147483647,
-                        "step": 1,
-                        "control_after_generate": True,
-                        "display": "number",
-                    },
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    control_after_generate=True,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    optional=True,
                 ),
-                "num_images": (
-                    IO.INT,
-                    {"default": 1, "min": 1, "max": 8, "step": 1, "display": "number"},
+                comfy_io.Int.Input(
+                    "num_images",
+                    default=1,
+                    min=1,
+                    max=8,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    optional=True,
                 ),
-                "rendering_speed": (
-                    IO.COMBO,
-                    {
-                        "options": ["BALANCED", "TURBO", "QUALITY"],
-                        "default": "BALANCED",
-                        "tooltip": "Controls the trade-off between generation speed and quality",
-                    },
+                comfy_io.Combo.Input(
+                    "rendering_speed",
+                    options=["BALANCED", "TURBO", "QUALITY"],
+                    default="BALANCED",
+                    tooltip="Controls the trade-off between generation speed and quality",
+                    optional=True,
                 ),
-            },
-            "hidden": {
-                "auth_token": "AUTH_TOKEN_COMFY_ORG",
-                "comfy_api_key": "API_KEY_COMFY_ORG",
-                "unique_id": "UNIQUE_ID",
-            },
-        }
+            ],
+            outputs=[
+                comfy_io.Image.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+        )
 
-    RETURN_TYPES = (IO.IMAGE,)
-    FUNCTION = "api_call"
-    CATEGORY = "api node/image/Ideogram"
-    DESCRIPTION = cleandoc(__doc__ or "")
-    API_NODE = True
-
-    async def api_call(
-        self,
+    @classmethod
+    async def execute(
+        cls,
         prompt,
         image=None,
         mask=None,
@@ -664,9 +640,11 @@ class IdeogramV3(ComfyNodeABC):
         seed=0,
         num_images=1,
         rendering_speed="BALANCED",
-        unique_id=None,
-        **kwargs,
     ):
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
         # Check if both image and mask are provided for editing mode
         if image is not None and mask is not None:
             # Edit mode
@@ -686,7 +664,7 @@ class IdeogramV3(ComfyNodeABC):
             # Process image
             img_np = (input_tensor.numpy() * 255).astype(np.uint8)
             img = Image.fromarray(img_np)
-            img_byte_arr = io.BytesIO()
+            img_byte_arr = BytesIO()
             img.save(img_byte_arr, format="PNG")
             img_byte_arr.seek(0)
             img_binary = img_byte_arr
@@ -695,7 +673,7 @@ class IdeogramV3(ComfyNodeABC):
             # Process mask - white areas will be replaced
             mask_np = (mask.squeeze().cpu().numpy() * 255).astype(np.uint8)
             mask_img = Image.fromarray(mask_np)
-            mask_byte_arr = io.BytesIO()
+            mask_byte_arr = BytesIO()
             mask_img.save(mask_byte_arr, format="PNG")
             mask_byte_arr.seek(0)
             mask_binary = mask_byte_arr
@@ -729,7 +707,7 @@ class IdeogramV3(ComfyNodeABC):
                     "mask": mask_binary,
                 },
                 content_type="multipart/form-data",
-                auth_kwargs=kwargs,
+                auth_kwargs=auth,
             )
 
         elif image is not None or mask is not None:
@@ -770,7 +748,7 @@ class IdeogramV3(ComfyNodeABC):
                     response_model=IdeogramGenerateResponse,
                 ),
                 request=gen_request,
-                auth_kwargs=kwargs,
+                auth_kwargs=auth,
             )
 
         # Execute the operation and process response
@@ -784,18 +762,18 @@ class IdeogramV3(ComfyNodeABC):
         if not image_urls:
             raise Exception("No image URLs were generated in the response")
 
-        display_image_urls_on_node(image_urls, unique_id)
-        return (await download_and_process_images(image_urls),)
+        display_image_urls_on_node(image_urls, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_and_process_images(image_urls))
 
 
-NODE_CLASS_MAPPINGS = {
-    "IdeogramV1": IdeogramV1,
-    "IdeogramV2": IdeogramV2,
-    "IdeogramV3": IdeogramV3,
-}
+class IdeogramExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[comfy_io.ComfyNode]]:
+        return [
+            IdeogramV1,
+            IdeogramV2,
+            IdeogramV3,
+        ]
 
-NODE_DISPLAY_NAME_MAPPINGS = {
-    "IdeogramV1": "Ideogram V1",
-    "IdeogramV2": "Ideogram V2",
-    "IdeogramV3": "Ideogram V3",
-}
+async def comfy_entrypoint() -> IdeogramExtension:
+    return IdeogramExtension()

comfy_api_nodes/nodes_kling.py

@@ -421,6 +421,8 @@ class KlingTextToVideoNode(KlingNodeBase):
             "pro mode / 10s duration / kling-v2-master": ("pro", "10", "kling-v2-master"),
             "standard mode / 5s duration / kling-v2-master": ("std", "5", "kling-v2-master"),
             "standard mode / 10s duration / kling-v2-master": ("std", "10", "kling-v2-master"),
+            "pro mode / 5s duration / kling-v2-1-master": ("pro", "5", "kling-v2-1-master"),
+            "pro mode / 10s duration / kling-v2-1-master": ("pro", "10", "kling-v2-1-master"),
         }
 
     @classmethod
@@ -1690,7 +1692,11 @@ class KlingImageGenerationNode(KlingImageGenerationBase):
     ):
         self.validate_prompt(prompt, negative_prompt)
 
-        if image is not None:
+        if image is None:
+            image_type = None
+        elif model_name == KlingImageGenModelName.kling_v1:
+            raise ValueError(f"The model {KlingImageGenModelName.kling_v1.value} does not support reference images.")
+        else:
             image = tensor_to_base64_string(image)
 
         initial_operation = SynchronousOperation(

comfy_api_nodes/nodes_minimax.py

@@ -1,3 +1,4 @@
+from inspect import cleandoc
 from typing import Union
 import logging
 import torch
@@ -10,7 +11,7 @@ from comfy_api_nodes.apis import (
     MinimaxFileRetrieveResponse,
     MinimaxTaskResultResponse,
     SubjectReferenceItem,
-    Model
+    MiniMaxModel
 )
 from comfy_api_nodes.apis.client import (
     ApiEndpoint,
@@ -84,7 +85,6 @@ class MinimaxTextToVideoNode:
     FUNCTION = "generate_video"
     CATEGORY = "api node/video/MiniMax"
     API_NODE = True
-    OUTPUT_NODE = True
 
     async def generate_video(
         self,
@@ -121,7 +121,7 @@ class MinimaxTextToVideoNode:
                 response_model=MinimaxVideoGenerationResponse,
             ),
             request=MinimaxVideoGenerationRequest(
-                model=Model(model),
+                model=MiniMaxModel(model),
                 prompt=prompt_text,
                 callback_url=None,
                 first_frame_image=image_url,
@@ -251,7 +251,6 @@ class MinimaxImageToVideoNode(MinimaxTextToVideoNode):
     FUNCTION = "generate_video"
     CATEGORY = "api node/video/MiniMax"
     API_NODE = True
-    OUTPUT_NODE = True
 
 
 class MinimaxSubjectToVideoNode(MinimaxTextToVideoNode):
@@ -313,7 +312,181 @@ class MinimaxSubjectToVideoNode(MinimaxTextToVideoNode):
     FUNCTION = "generate_video"
     CATEGORY = "api node/video/MiniMax"
     API_NODE = True
-    OUTPUT_NODE = True
+
+
+class MinimaxHailuoVideoNode:
+    """Generates videos from prompt, with optional start frame using the new MiniMax Hailuo-02 model."""
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "prompt_text": (
+                    "STRING",
+                    {
+                        "multiline": True,
+                        "default": "",
+                        "tooltip": "Text prompt to guide the video generation.",
+                    },
+                ),
+            },
+            "optional": {
+                "seed": (
+                    IO.INT,
+                    {
+                        "default": 0,
+                        "min": 0,
+                        "max": 0xFFFFFFFFFFFFFFFF,
+                        "control_after_generate": True,
+                        "tooltip": "The random seed used for creating the noise.",
+                    },
+                ),
+                "first_frame_image": (
+                    IO.IMAGE,
+                    {
+                        "tooltip": "Optional image to use as the first frame to generate a video."
+                    },
+                ),
+                "prompt_optimizer": (
+                    IO.BOOLEAN,
+                    {
+                        "tooltip": "Optimize prompt to improve generation quality when needed.",
+                        "default": True,
+                    },
+                ),
+                "duration": (
+                    IO.COMBO,
+                    {
+                        "tooltip": "The length of the output video in seconds.",
+                        "default": 6,
+                        "options": [6, 10],
+                    },
+                ),
+                "resolution": (
+                    IO.COMBO,
+                    {
+                        "tooltip": "The dimensions of the video display. "
+                                   "1080p corresponds to 1920 x 1080 pixels, 768p corresponds to 1366 x 768 pixels.",
+                        "default": "768P",
+                        "options": ["768P", "1080P"],
+                    },
+                ),
+            },
+            "hidden": {
+                "auth_token": "AUTH_TOKEN_COMFY_ORG",
+                "comfy_api_key": "API_KEY_COMFY_ORG",
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+
+    RETURN_TYPES = ("VIDEO",)
+    DESCRIPTION = cleandoc(__doc__ or "")
+    FUNCTION = "generate_video"
+    CATEGORY = "api node/video/MiniMax"
+    API_NODE = True
+
+    async def generate_video(
+        self,
+        prompt_text,
+        seed=0,
+        first_frame_image: torch.Tensor=None, # used for ImageToVideo
+        prompt_optimizer=True,
+        duration=6,
+        resolution="768P",
+        model="MiniMax-Hailuo-02",
+        unique_id: Union[str, None]=None,
+        **kwargs,
+    ):
+        if first_frame_image is None:
+            validate_string(prompt_text, field_name="prompt_text")
+
+        if model == "MiniMax-Hailuo-02" and resolution.upper() == "1080P" and duration != 6:
+            raise Exception(
+                "When model is MiniMax-Hailuo-02 and resolution is 1080P, duration is limited to 6 seconds."
+            )
+
+        # upload image, if passed in
+        image_url = None
+        if first_frame_image is not None:
+            image_url = (await upload_images_to_comfyapi(first_frame_image, max_images=1, auth_kwargs=kwargs))[0]
+
+        video_generate_operation = SynchronousOperation(
+            endpoint=ApiEndpoint(
+                path="/proxy/minimax/video_generation",
+                method=HttpMethod.POST,
+                request_model=MinimaxVideoGenerationRequest,
+                response_model=MinimaxVideoGenerationResponse,
+            ),
+            request=MinimaxVideoGenerationRequest(
+                model=MiniMaxModel(model),
+                prompt=prompt_text,
+                callback_url=None,
+                first_frame_image=image_url,
+                prompt_optimizer=prompt_optimizer,
+                duration=duration,
+                resolution=resolution,
+            ),
+            auth_kwargs=kwargs,
+        )
+        response = await video_generate_operation.execute()
+
+        task_id = response.task_id
+        if not task_id:
+            raise Exception(f"MiniMax generation failed: {response.base_resp}")
+
+        average_duration = 120 if resolution == "768P" else 240
+        video_generate_operation = PollingOperation(
+            poll_endpoint=ApiEndpoint(
+                path="/proxy/minimax/query/video_generation",
+                method=HttpMethod.GET,
+                request_model=EmptyRequest,
+                response_model=MinimaxTaskResultResponse,
+                query_params={"task_id": task_id},
+            ),
+            completed_statuses=["Success"],
+            failed_statuses=["Fail"],
+            status_extractor=lambda x: x.status.value,
+            estimated_duration=average_duration,
+            node_id=unique_id,
+            auth_kwargs=kwargs,
+        )
+        task_result = await video_generate_operation.execute()
+
+        file_id = task_result.file_id
+        if file_id is None:
+            raise Exception("Request was not successful. Missing file ID.")
+        file_retrieve_operation = SynchronousOperation(
+            endpoint=ApiEndpoint(
+                path="/proxy/minimax/files/retrieve",
+                method=HttpMethod.GET,
+                request_model=EmptyRequest,
+                response_model=MinimaxFileRetrieveResponse,
+                query_params={"file_id": int(file_id)},
+            ),
+            request=EmptyRequest(),
+            auth_kwargs=kwargs,
+        )
+        file_result = await file_retrieve_operation.execute()
+
+        file_url = file_result.file.download_url
+        if file_url is None:
+            raise Exception(
+                f"No video was found in the response. Full response: {file_result.model_dump()}"
+            )
+        logging.info(f"Generated video URL: {file_url}")
+        if unique_id:
+            if hasattr(file_result.file, "backup_download_url"):
+                message = f"Result URL: {file_url}\nBackup URL: {file_result.file.backup_download_url}"
+            else:
+                message = f"Result URL: {file_url}"
+            PromptServer.instance.send_progress_text(message, unique_id)
+
+        video_io = await download_url_to_bytesio(file_url)
+        if video_io is None:
+            error_msg = f"Failed to download video from {file_url}"
+            logging.error(error_msg)
+            raise Exception(error_msg)
+        return (VideoFromFile(video_io),)
 
 
 # A dictionary that contains all nodes you want to export with their names
@@ -322,6 +495,7 @@ NODE_CLASS_MAPPINGS = {
     "MinimaxTextToVideoNode": MinimaxTextToVideoNode,
     "MinimaxImageToVideoNode": MinimaxImageToVideoNode,
     # "MinimaxSubjectToVideoNode": MinimaxSubjectToVideoNode,
+    "MinimaxHailuoVideoNode": MinimaxHailuoVideoNode,
 }
 
 # A dictionary that contains the friendly/humanly readable titles for the nodes
@@ -329,4 +503,5 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "MinimaxTextToVideoNode": "MiniMax Text to Video",
     "MinimaxImageToVideoNode": "MiniMax Image to Video",
     "MinimaxSubjectToVideoNode": "MiniMax Subject to Video",
+    "MinimaxHailuoVideoNode": "MiniMax Hailuo Video",
 }

comfy_api_nodes/nodes_moonvalley.py

@@ -1,6 +1,5 @@
 import logging
 from typing import Any, Callable, Optional, TypeVar
-import random
 import torch
 from comfy_api_nodes.util.validation_utils import (
     get_image_dimensions,
@@ -208,20 +207,29 @@ def _get_video_dimensions(video: VideoInput) -> tuple[int, int]:
 def _validate_video_dimensions(width: int, height: int) -> None:
     """Validates video dimensions meet Moonvalley V2V requirements."""
     supported_resolutions = {
-        (1920, 1080), (1080, 1920), (1152, 1152),
-        (1536, 1152), (1152, 1536)
+        (1920, 1080),
+        (1080, 1920),
+        (1152, 1152),
+        (1536, 1152),
+        (1152, 1536),
     }
 
     if (width, height) not in supported_resolutions:
-        supported_list = ', '.join([f'{w}x{h}' for w, h in sorted(supported_resolutions)])
-        raise ValueError(f"Resolution {width}x{height} not supported. Supported: {supported_list}")
+        supported_list = ", ".join(
+            [f"{w}x{h}" for w, h in sorted(supported_resolutions)]
+        )
+        raise ValueError(
+            f"Resolution {width}x{height} not supported. Supported: {supported_list}"
+        )
 
 
 def _validate_container_format(video: VideoInput) -> None:
     """Validates video container format is MP4."""
     container_format = video.get_container_format()
-    if container_format not in ['mp4', 'mov,mp4,m4a,3gp,3g2,mj2']:
-        raise ValueError(f"Only MP4 container format supported. Got: {container_format}")
+    if container_format not in ["mp4", "mov,mp4,m4a,3gp,3g2,mj2"]:
+        raise ValueError(
+            f"Only MP4 container format supported. Got: {container_format}"
+        )
 
 
 def _validate_and_trim_duration(video: VideoInput) -> VideoInput:
@@ -244,7 +252,6 @@ def _trim_if_too_long(video: VideoInput, duration: float) -> VideoInput:
     return video
 
 
-
 def trim_video(video: VideoInput, duration_sec: float) -> VideoInput:
     """
     Returns a new VideoInput object trimmed from the beginning to the specified duration,
@@ -302,7 +309,9 @@ def trim_video(video: VideoInput, duration_sec: float) -> VideoInput:
         # Calculate target frame count that's divisible by 16
         fps = input_container.streams.video[0].average_rate
         estimated_frames = int(duration_sec * fps)
-        target_frames = (estimated_frames // 16) * 16  # Round down to nearest multiple of 16
+        target_frames = (
+            estimated_frames // 16
+        ) * 16  # Round down to nearest multiple of 16
 
         if target_frames == 0:
             raise ValueError("Video too short: need at least 16 frames for Moonvalley")
@@ -424,7 +433,7 @@ class BaseMoonvalleyVideoNode:
                     MoonvalleyTextToVideoInferenceParams,
                     "negative_prompt",
                     multiline=True,
-                    default="low-poly, flat shader, bad rigging, stiff animation, uncanny eyes, low-quality textures, looping glitch, cheap effect, overbloom, bloom spam, default lighting, game asset, stiff face, ugly specular, AI artifacts",
+                    default="<synthetic> <scene cut> gopro, bright, contrast, static, overexposed, vignette, artifacts, still, noise, texture, scanlines, videogame, 360 camera, VR, transition, flare, saturation, distorted, warped, wide angle, saturated, vibrant, glowing, cross dissolve, cheesy, ugly hands, mutated hands, mutant, disfigured, extra fingers, blown out, horrible, blurry, worst quality, bad, dissolve, melt, fade in, fade out, wobbly, weird, low quality, plastic, stock footage, video camera, boring",
                 ),
                 "resolution": (
                     IO.COMBO,
@@ -441,12 +450,11 @@ class BaseMoonvalleyVideoNode:
                         "tooltip": "Resolution of the output video",
                     },
                 ),
-                # "length": (IO.COMBO,{"options":['5s','10s'], "default": '5s'}),
                 "prompt_adherence": model_field_to_node_input(
                     IO.FLOAT,
                     MoonvalleyTextToVideoInferenceParams,
                     "guidance_scale",
-                    default=7.0,
+                    default=10.0,
                     step=1,
                     min=1,
                     max=20,
@@ -455,13 +463,12 @@ class BaseMoonvalleyVideoNode:
                     IO.INT,
                     MoonvalleyTextToVideoInferenceParams,
                     "seed",
-                    default=random.randint(0, 2**32 - 1),
+                    default=9,
                     min=0,
                     max=4294967295,
                     step=1,
                     display="number",
                     tooltip="Random seed value",
-                    control_after_generate=True,
                 ),
                 "steps": model_field_to_node_input(
                     IO.INT,
@@ -532,9 +539,11 @@ class MoonvalleyImg2VideoNode(BaseMoonvalleyVideoNode):
         # Get MIME type from tensor - assuming PNG format for image tensors
         mime_type = "image/png"
 
-        image_url = (await upload_images_to_comfyapi(
-            image, max_images=1, auth_kwargs=kwargs, mime_type=mime_type
-        ))[0]
+        image_url = (
+            await upload_images_to_comfyapi(
+                image, max_images=1, auth_kwargs=kwargs, mime_type=mime_type
+            )
+        )[0]
 
         request = MoonvalleyTextToVideoRequest(
             image_url=image_url, prompt_text=prompt, inference_params=inference_params
@@ -570,17 +579,39 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
         return {
             "required": {
                 "prompt": model_field_to_node_input(
-                    IO.STRING, MoonvalleyVideoToVideoRequest, "prompt_text",
-                    multiline=True
+                    IO.STRING,
+                    MoonvalleyVideoToVideoRequest,
+                    "prompt_text",
+                    multiline=True,
                 ),
                 "negative_prompt": model_field_to_node_input(
                     IO.STRING,
                     MoonvalleyVideoToVideoInferenceParams,
                     "negative_prompt",
                     multiline=True,
-                    default="low-poly, flat shader, bad rigging, stiff animation, uncanny eyes, low-quality textures, looping glitch, cheap effect, overbloom, bloom spam, default lighting, game asset, stiff face, ugly specular, AI artifacts"
+                    default="<synthetic> <scene cut> gopro, bright, contrast, static, overexposed, vignette, artifacts, still, noise, texture, scanlines, videogame, 360 camera, VR, transition, flare, saturation, distorted, warped, wide angle, saturated, vibrant, glowing, cross dissolve, cheesy, ugly hands, mutated hands, mutant, disfigured, extra fingers, blown out, horrible, blurry, worst quality, bad, dissolve, melt, fade in, fade out, wobbly, weird, low quality, plastic, stock footage, video camera, boring",
+                ),
+                "seed": model_field_to_node_input(
+                    IO.INT,
+                    MoonvalleyVideoToVideoInferenceParams,
+                    "seed",
+                    default=9,
+                    min=0,
+                    max=4294967295,
+                    step=1,
+                    display="number",
+                    tooltip="Random seed value",
+                    control_after_generate=False,
+                ),
+                "prompt_adherence": model_field_to_node_input(
+                    IO.FLOAT,
+                    MoonvalleyVideoToVideoInferenceParams,
+                    "guidance_scale",
+                    default=10.0,
+                    step=1,
+                    min=1,
+                    max=20,
                 ),
-                "seed": model_field_to_node_input(IO.INT,MoonvalleyVideoToVideoInferenceParams, "seed", default=random.randint(0, 2**32 - 1), min=0, max=4294967295, step=1, display="number", tooltip="Random seed value", control_after_generate=True),
             },
             "hidden": {
                 "auth_token": "AUTH_TOKEN_COMFY_ORG",
@@ -588,7 +619,14 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
                 "unique_id": "UNIQUE_ID",
             },
             "optional": {
-                "video": (IO.VIDEO, {"default": "", "multiline": False, "tooltip": "The reference video used to generate the output video. Must be at least 5 seconds long. Videos longer than 5s will be automatically trimmed. Only MP4 format supported."}),
+                "video": (
+                    IO.VIDEO,
+                    {
+                        "default": "",
+                        "multiline": False,
+                        "tooltip": "The reference video used to generate the output video. Must be at least 5 seconds long. Videos longer than 5s will be automatically trimmed. Only MP4 format supported.",
+                    },
+                ),
                 "control_type": (
                     ["Motion Transfer", "Pose Transfer"],
                     {"default": "Motion Transfer"},
@@ -602,8 +640,14 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
                         "max": 100,
                         "tooltip": "Only used if control_type is 'Motion Transfer'",
                     },
-                )
-            }
+                ),
+                "image": model_field_to_node_input(
+                    IO.IMAGE,
+                    MoonvalleyTextToVideoRequest,
+                    "image_url",
+                    tooltip="The reference image used to generate the video",
+                ),
+            },
         }
 
     RETURN_TYPES = ("VIDEO",)
@@ -613,6 +657,7 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
         self, prompt, negative_prompt, unique_id: Optional[str] = None, **kwargs
     ):
         video = kwargs.get("video")
+        image = kwargs.get("image", None)
 
         if not video:
             raise MoonvalleyApiError("video is required")
@@ -620,8 +665,16 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
         video_url = ""
         if video:
             validated_video = validate_video_to_video_input(video)
-            video_url = await upload_video_to_comfyapi(validated_video, auth_kwargs=kwargs)
+            video_url = await upload_video_to_comfyapi(
+                validated_video, auth_kwargs=kwargs
+            )
+        mime_type = "image/png"
 
+        if not image is None:
+            validate_input_image(image, with_frame_conditioning=True)
+            image_url = await upload_images_to_comfyapi(
+                image=image, auth_kwargs=kwargs, max_images=1, mime_type=mime_type
+            )
         control_type = kwargs.get("control_type")
         motion_intensity = kwargs.get("motion_intensity")
 
@@ -631,12 +684,12 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
         # Only include motion_intensity for Motion Transfer
         control_params = {}
         if control_type == "Motion Transfer" and motion_intensity is not None:
-            control_params['motion_intensity'] = motion_intensity
+            control_params["motion_intensity"] = motion_intensity
 
-        inference_params=MoonvalleyVideoToVideoInferenceParams(
+        inference_params = MoonvalleyVideoToVideoInferenceParams(
             negative_prompt=negative_prompt,
             seed=kwargs.get("seed"),
-            control_params=control_params
+            control_params=control_params,
         )
 
         control = self.parseControlParameter(control_type)
@@ -647,6 +700,7 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
             prompt_text=prompt,
             inference_params=inference_params,
         )
+        request.image_url = image_url if not image is None else None
 
         initial_operation = SynchronousOperation(
             endpoint=ApiEndpoint(
@@ -694,15 +748,15 @@ class MoonvalleyTxt2VideoNode(BaseMoonvalleyVideoNode):
         validate_prompts(prompt, negative_prompt, MOONVALLEY_MAREY_MAX_PROMPT_LENGTH)
         width_height = self.parseWidthHeightFromRes(kwargs.get("resolution"))
 
-        inference_params=MoonvalleyTextToVideoInferenceParams(
-                    negative_prompt=negative_prompt,
-                    steps=kwargs.get("steps"),
-                    seed=kwargs.get("seed"),
-                    guidance_scale=kwargs.get("prompt_adherence"),
-                    num_frames=128,
-                    width=width_height.get("width"),
-                    height=width_height.get("height"),
-                )
+        inference_params = MoonvalleyTextToVideoInferenceParams(
+            negative_prompt=negative_prompt,
+            steps=kwargs.get("steps"),
+            seed=kwargs.get("seed"),
+            guidance_scale=kwargs.get("prompt_adherence"),
+            num_frames=128,
+            width=width_height.get("width"),
+            height=width_height.get("height"),
+        )
         request = MoonvalleyTextToVideoRequest(
             prompt_text=prompt, inference_params=inference_params
         )

comfy_api_nodes/nodes_openai.py

@@ -80,6 +80,9 @@ class SupportedOpenAIModel(str, Enum):
     gpt_4_1 = "gpt-4.1"
     gpt_4_1_mini = "gpt-4.1-mini"
     gpt_4_1_nano = "gpt-4.1-nano"
+    gpt_5 = "gpt-5"
+    gpt_5_mini = "gpt-5-mini"
+    gpt_5_nano = "gpt-5-nano"
 
 
 class OpenAIDalle2(ComfyNodeABC):
@@ -464,8 +467,6 @@ class OpenAIGPTImage1(ComfyNodeABC):
         path = "/proxy/openai/images/generations"
         content_type = "application/json"
         request_class = OpenAIImageGenerationRequest
-        img_binaries = []
-        mask_binary = None
         files = []
 
         if image is not None:
@@ -484,14 +485,11 @@ class OpenAIGPTImage1(ComfyNodeABC):
                 img_byte_arr = io.BytesIO()
                 img.save(img_byte_arr, format="PNG")
                 img_byte_arr.seek(0)
-                img_binary = img_byte_arr
-                img_binary.name = f"image_{i}.png"
 
-                img_binaries.append(img_binary)
                 if batch_size == 1:
-                    files.append(("image", img_binary))
+                    files.append(("image", (f"image_{i}.png", img_byte_arr, "image/png")))
                 else:
-                    files.append(("image[]", img_binary))
+                    files.append(("image[]", (f"image_{i}.png", img_byte_arr, "image/png")))
 
         if mask is not None:
             if image is None:
@@ -511,9 +509,7 @@ class OpenAIGPTImage1(ComfyNodeABC):
             mask_img_byte_arr = io.BytesIO()
             mask_img.save(mask_img_byte_arr, format="PNG")
             mask_img_byte_arr.seek(0)
-            mask_binary = mask_img_byte_arr
-            mask_binary.name = "mask.png"
-            files.append(("mask", mask_binary))
+            files.append(("mask", ("mask.png", mask_img_byte_arr, "image/png")))
 
         # Build the operation
         operation = SynchronousOperation(
@@ -1002,7 +998,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "OpenAIDalle2": "OpenAI DALL·E 2",
     "OpenAIDalle3": "OpenAI DALL·E 3",
     "OpenAIGPTImage1": "OpenAI GPT Image 1",
-    "OpenAIChatNode": "OpenAI Chat",
-    "OpenAIInputFiles": "OpenAI Chat Input Files",
-    "OpenAIChatConfig": "OpenAI Chat Advanced Options",
+    "OpenAIChatNode": "OpenAI ChatGPT",
+    "OpenAIInputFiles": "OpenAI ChatGPT Input Files",
+    "OpenAIChatConfig": "OpenAI ChatGPT Advanced Options",
 }

comfy_api_nodes/nodes_veo2.py

@@ -1,17 +1,18 @@
-import io
 import logging
 import base64
 import aiohttp
 import torch
+from io import BytesIO
 from typing import Optional
+from typing_extensions import override
 
-from comfy.comfy_types.node_typing import IO, ComfyNodeABC
+from comfy_api.latest import ComfyExtension, io as comfy_io
 from comfy_api.input_impl.video_types import VideoFromFile
 from comfy_api_nodes.apis import (
     VeoGenVidRequest,
     VeoGenVidResponse,
     VeoGenVidPollRequest,
-    VeoGenVidPollResponse
+    VeoGenVidPollResponse,
 )
 from comfy_api_nodes.apis.client import (
     ApiEndpoint,
@@ -22,7 +23,7 @@ from comfy_api_nodes.apis.client import (
 
 from comfy_api_nodes.apinode_utils import (
     downscale_image_tensor,
-    tensor_to_base64_string
+    tensor_to_base64_string,
 )
 
 AVERAGE_DURATION_VIDEO_GEN = 32
@@ -50,7 +51,7 @@ def get_video_url_from_response(poll_response: VeoGenVidPollResponse) -> Optiona
     return None
 
 
-class VeoVideoGenerationNode(ComfyNodeABC):
+class VeoVideoGenerationNode(comfy_io.ComfyNode):
     """
     Generates videos from text prompts using Google's Veo API.
 
@@ -59,101 +60,93 @@ class VeoVideoGenerationNode(ComfyNodeABC):
     """
 
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Text description of the video",
-                    },
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="VeoVideoGenerationNode",
+            display_name="Google Veo 2 Video Generation",
+            category="api node/video/Veo",
+            description="Generates videos from text prompts using Google's Veo 2 API",
+            inputs=[
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Text description of the video",
                 ),
-                "aspect_ratio": (
-                    IO.COMBO,
-                    {
-                        "options": ["16:9", "9:16"],
-                        "default": "16:9",
-                        "tooltip": "Aspect ratio of the output video",
-                    },
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=["16:9", "9:16"],
+                    default="16:9",
+                    tooltip="Aspect ratio of the output video",
                 ),
-            },
-            "optional": {
-                "negative_prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Negative text prompt to guide what to avoid in the video",
-                    },
+                comfy_io.String.Input(
+                    "negative_prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Negative text prompt to guide what to avoid in the video",
+                    optional=True,
                 ),
-                "duration_seconds": (
-                    IO.INT,
-                    {
-                        "default": 5,
-                        "min": 5,
-                        "max": 8,
-                        "step": 1,
-                        "display": "number",
-                        "tooltip": "Duration of the output video in seconds",
-                    },
+                comfy_io.Int.Input(
+                    "duration_seconds",
+                    default=5,
+                    min=5,
+                    max=8,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
                 ),
-                "enhance_prompt": (
-                    IO.BOOLEAN,
-                    {
-                        "default": True,
-                        "tooltip": "Whether to enhance the prompt with AI assistance",
-                    }
+                comfy_io.Boolean.Input(
+                    "enhance_prompt",
+                    default=True,
+                    tooltip="Whether to enhance the prompt with AI assistance",
+                    optional=True,
                 ),
-                "person_generation": (
-                    IO.COMBO,
-                    {
-                        "options": ["ALLOW", "BLOCK"],
-                        "default": "ALLOW",
-                        "tooltip": "Whether to allow generating people in the video",
-                    },
+                comfy_io.Combo.Input(
+                    "person_generation",
+                    options=["ALLOW", "BLOCK"],
+                    default="ALLOW",
+                    tooltip="Whether to allow generating people in the video",
+                    optional=True,
                 ),
-                "seed": (
-                    IO.INT,
-                    {
-                        "default": 0,
-                        "min": 0,
-                        "max": 0xFFFFFFFF,
-                        "step": 1,
-                        "display": "number",
-                        "control_after_generate": True,
-                        "tooltip": "Seed for video generation (0 for random)",
-                    },
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=0xFFFFFFFF,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
                 ),
-                "image": (IO.IMAGE, {
-                    "default": None,
-                    "tooltip": "Optional reference image to guide video generation",
-                }),
-                "model": (
-                    IO.COMBO,
-                    {
-                        "options": ["veo-2.0-generate-001"],
-                        "default": "veo-2.0-generate-001",
-                        "tooltip": "Veo 2 model to use for video generation",
-                    },
+                comfy_io.Image.Input(
+                    "image",
+                    tooltip="Optional reference image to guide video generation",
+                    optional=True,
                 ),
-            },
-            "hidden": {
-                "auth_token": "AUTH_TOKEN_COMFY_ORG",
-                "comfy_api_key": "API_KEY_COMFY_ORG",
-                "unique_id": "UNIQUE_ID",
-            },
-        }
+                comfy_io.Combo.Input(
+                    "model",
+                    options=["veo-2.0-generate-001"],
+                    default="veo-2.0-generate-001",
+                    tooltip="Veo 2 model to use for video generation",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
 
-    RETURN_TYPES = (IO.VIDEO,)
-    FUNCTION = "generate_video"
-    CATEGORY = "api node/video/Veo"
-    DESCRIPTION = "Generates videos from text prompts using Google's Veo 2 API"
-    API_NODE = True
-
-    async def generate_video(
-        self,
+    @classmethod
+    async def execute(
+        cls,
         prompt,
         aspect_ratio="16:9",
         negative_prompt="",
@@ -164,8 +157,6 @@ class VeoVideoGenerationNode(ComfyNodeABC):
         image=None,
         model="veo-2.0-generate-001",
         generate_audio=False,
-        unique_id: Optional[str] = None,
-        **kwargs,
     ):
         # Prepare the instances for the request
         instances = []
@@ -202,6 +193,10 @@ class VeoVideoGenerationNode(ComfyNodeABC):
         if "veo-3.0" in model:
             parameters["generateAudio"] = generate_audio
 
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
         # Initial request to start video generation
         initial_operation = SynchronousOperation(
             endpoint=ApiEndpoint(
@@ -214,7 +209,7 @@ class VeoVideoGenerationNode(ComfyNodeABC):
                 instances=instances,
                 parameters=parameters
             ),
-            auth_kwargs=kwargs,
+            auth_kwargs=auth,
         )
 
         initial_response = await initial_operation.execute()
@@ -248,10 +243,10 @@ class VeoVideoGenerationNode(ComfyNodeABC):
             request=VeoGenVidPollRequest(
                 operationName=operation_name
             ),
-            auth_kwargs=kwargs,
+            auth_kwargs=auth,
             poll_interval=5.0,
             result_url_extractor=get_video_url_from_response,
-            node_id=unique_id,
+            node_id=cls.hidden.unique_id,
             estimated_duration=AVERAGE_DURATION_VIDEO_GEN,
         )
 
@@ -304,10 +299,10 @@ class VeoVideoGenerationNode(ComfyNodeABC):
         logging.info("Video generation completed successfully")
 
         # Convert video data to BytesIO object
-        video_io = io.BytesIO(video_data)
+        video_io = BytesIO(video_data)
 
         # Return VideoFromFile object
-        return (VideoFromFile(video_io),)
+        return comfy_io.NodeOutput(VideoFromFile(video_io))
 
 
 class Veo3VideoGenerationNode(VeoVideoGenerationNode):
@@ -323,51 +318,104 @@ class Veo3VideoGenerationNode(VeoVideoGenerationNode):
     """
 
     @classmethod
-    def INPUT_TYPES(s):
-        parent_input = super().INPUT_TYPES()
-
-        # Update model options for Veo 3
-        parent_input["optional"]["model"] = (
-            IO.COMBO,
-            {
-                "options": ["veo-3.0-generate-001", "veo-3.0-fast-generate-001"],
-                "default": "veo-3.0-generate-001",
-                "tooltip": "Veo 3 model to use for video generation",
-            },
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="Veo3VideoGenerationNode",
+            display_name="Google Veo 3 Video Generation",
+            category="api node/video/Veo",
+            description="Generates videos from text prompts using Google's Veo 3 API",
+            inputs=[
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Text description of the video",
+                ),
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=["16:9", "9:16"],
+                    default="16:9",
+                    tooltip="Aspect ratio of the output video",
+                ),
+                comfy_io.String.Input(
+                    "negative_prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Negative text prompt to guide what to avoid in the video",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "duration_seconds",
+                    default=8,
+                    min=8,
+                    max=8,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds (Veo 3 only supports 8 seconds)",
+                    optional=True,
+                ),
+                comfy_io.Boolean.Input(
+                    "enhance_prompt",
+                    default=True,
+                    tooltip="Whether to enhance the prompt with AI assistance",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "person_generation",
+                    options=["ALLOW", "BLOCK"],
+                    default="ALLOW",
+                    tooltip="Whether to allow generating people in the video",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=0xFFFFFFFF,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Image.Input(
+                    "image",
+                    tooltip="Optional reference image to guide video generation",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "model",
+                    options=["veo-3.0-generate-001", "veo-3.0-fast-generate-001"],
+                    default="veo-3.0-generate-001",
+                    tooltip="Veo 3 model to use for video generation",
+                    optional=True,
+                ),
+                comfy_io.Boolean.Input(
+                    "generate_audio",
+                    default=False,
+                    tooltip="Generate audio for the video. Supported by all Veo 3 models.",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
         )
 
-        # Add generateAudio parameter
-        parent_input["optional"]["generate_audio"] = (
-            IO.BOOLEAN,
-            {
-                "default": False,
-                "tooltip": "Generate audio for the video. Supported by all Veo 3 models.",
-            }
-        )
 
-        # Update duration constraints for Veo 3 (only 8 seconds supported)
-        parent_input["optional"]["duration_seconds"] = (
-            IO.INT,
-            {
-                "default": 8,
-                "min": 8,
-                "max": 8,
-                "step": 1,
-                "display": "number",
-                "tooltip": "Duration of the output video in seconds (Veo 3 only supports 8 seconds)",
-            },
-        )
+class VeoExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[comfy_io.ComfyNode]]:
+        return [
+            VeoVideoGenerationNode,
+            Veo3VideoGenerationNode,
+        ]
 
-        return parent_input
-
-
-# Register the nodes
-NODE_CLASS_MAPPINGS = {
-    "VeoVideoGenerationNode": VeoVideoGenerationNode,
-    "Veo3VideoGenerationNode": Veo3VideoGenerationNode,
-}
-
-NODE_DISPLAY_NAME_MAPPINGS = {
-    "VeoVideoGenerationNode": "Google Veo 2 Video Generation",
-    "Veo3VideoGenerationNode": "Google Veo 3 Video Generation",
-}
+async def comfy_entrypoint() -> VeoExtension:
+    return VeoExtension()

comfy_api_nodes/nodes_vidu.py

@@ -0,0 +1,622 @@
+import logging
+from enum import Enum
+from typing import Any, Callable, Optional, Literal, TypeVar
+from typing_extensions import override
+
+import torch
+from pydantic import BaseModel, Field
+
+from comfy_api.latest import ComfyExtension, io as comfy_io
+from comfy_api_nodes.util.validation_utils import (
+    validate_aspect_ratio_closeness,
+    validate_image_dimensions,
+    validate_image_aspect_ratio_range,
+    get_number_of_images,
+)
+from comfy_api_nodes.apis.client import (
+    ApiEndpoint,
+    HttpMethod,
+    SynchronousOperation,
+    PollingOperation,
+    EmptyRequest,
+)
+from comfy_api_nodes.apinode_utils import download_url_to_video_output, upload_images_to_comfyapi
+
+
+VIDU_TEXT_TO_VIDEO = "/proxy/vidu/text2video"
+VIDU_IMAGE_TO_VIDEO = "/proxy/vidu/img2video"
+VIDU_REFERENCE_VIDEO = "/proxy/vidu/reference2video"
+VIDU_START_END_VIDEO = "/proxy/vidu/start-end2video"
+VIDU_GET_GENERATION_STATUS = "/proxy/vidu/tasks/%s/creations"
+
+R = TypeVar("R")
+
+class VideoModelName(str, Enum):
+    vidu_q1 = 'viduq1'
+
+
+class AspectRatio(str, Enum):
+    r_16_9 = "16:9"
+    r_9_16 = "9:16"
+    r_1_1 = "1:1"
+
+
+class Resolution(str, Enum):
+    r_1080p = "1080p"
+
+
+class MovementAmplitude(str, Enum):
+    auto = "auto"
+    small = "small"
+    medium = "medium"
+    large = "large"
+
+
+class TaskCreationRequest(BaseModel):
+    model: VideoModelName = VideoModelName.vidu_q1
+    prompt: Optional[str] = Field(None, max_length=1500)
+    duration: Optional[Literal[5]] = 5
+    seed: Optional[int] = Field(0, ge=0, le=2147483647)
+    aspect_ratio: Optional[AspectRatio] = AspectRatio.r_16_9
+    resolution: Optional[Resolution] = Resolution.r_1080p
+    movement_amplitude: Optional[MovementAmplitude] = MovementAmplitude.auto
+    images: Optional[list[str]] = Field(None, description="Base64 encoded string or image URL")
+
+
+class TaskStatus(str, Enum):
+    created = "created"
+    queueing = "queueing"
+    processing = "processing"
+    success = "success"
+    failed = "failed"
+
+
+class TaskCreationResponse(BaseModel):
+    task_id: str = Field(...)
+    state: TaskStatus = Field(...)
+    created_at: str = Field(...)
+    code: Optional[int] = Field(None, description="Error code")
+
+
+class TaskResult(BaseModel):
+    id: str = Field(..., description="Creation id")
+    url: str = Field(..., description="The URL of the generated results, valid for one hour")
+    cover_url: str = Field(..., description="The cover URL of the generated results, valid for one hour")
+
+
+class TaskStatusResponse(BaseModel):
+    state: TaskStatus = Field(...)
+    err_code: Optional[str] = Field(None)
+    creations: list[TaskResult] = Field(..., description="Generated results")
+
+
+async def poll_until_finished(
+    auth_kwargs: dict[str, str],
+    api_endpoint: ApiEndpoint[Any, R],
+    result_url_extractor: Optional[Callable[[R], str]] = None,
+    estimated_duration: Optional[int] = None,
+    node_id: Optional[str] = None,
+) -> R:
+    return await PollingOperation(
+        poll_endpoint=api_endpoint,
+        completed_statuses=[TaskStatus.success.value],
+        failed_statuses=[TaskStatus.failed.value],
+        status_extractor=lambda response: response.state.value,
+        auth_kwargs=auth_kwargs,
+        result_url_extractor=result_url_extractor,
+        estimated_duration=estimated_duration,
+        node_id=node_id,
+        poll_interval=16.0,
+        max_poll_attempts=256,
+    ).execute()
+
+
+def get_video_url_from_response(response) -> Optional[str]:
+    if response.creations:
+        return response.creations[0].url
+    return None
+
+
+def get_video_from_response(response) -> TaskResult:
+    if not response.creations:
+        error_msg = f"Vidu request does not contain results. State: {response.state}, Error Code: {response.err_code}"
+        logging.info(error_msg)
+        raise RuntimeError(error_msg)
+    logging.info("Vidu task %s succeeded. Video URL: %s", response.creations[0].id, response.creations[0].url)
+    return response.creations[0]
+
+
+async def execute_task(
+    vidu_endpoint: str,
+    auth_kwargs: Optional[dict[str, str]],
+    payload: TaskCreationRequest,
+    estimated_duration: int,
+    node_id: str,
+) -> R:
+    response = await SynchronousOperation(
+        endpoint=ApiEndpoint(
+            path=vidu_endpoint,
+            method=HttpMethod.POST,
+            request_model=TaskCreationRequest,
+            response_model=TaskCreationResponse,
+        ),
+        request=payload,
+        auth_kwargs=auth_kwargs,
+    ).execute()
+    if response.state == TaskStatus.failed:
+        error_msg = f"Vidu request failed. Code: {response.code}"
+        logging.error(error_msg)
+        raise RuntimeError(error_msg)
+    return await poll_until_finished(
+        auth_kwargs,
+        ApiEndpoint(
+            path=VIDU_GET_GENERATION_STATUS % response.task_id,
+            method=HttpMethod.GET,
+            request_model=EmptyRequest,
+            response_model=TaskStatusResponse,
+        ),
+        result_url_extractor=get_video_url_from_response,
+        estimated_duration=estimated_duration,
+        node_id=node_id,
+    )
+
+
+class ViduTextToVideoNode(comfy_io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="ViduTextToVideoNode",
+            display_name="Vidu Text To Video Generation",
+            category="api node/video/Vidu",
+            description="Generate video from text prompt",
+            inputs=[
+                comfy_io.Combo.Input(
+                    "model",
+                    options=[model.value for model in VideoModelName],
+                    default=VideoModelName.vidu_q1.value,
+                    tooltip="Model name",
+                ),
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="A textual description for video generation",
+                ),
+                comfy_io.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=5,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=[model.value for model in AspectRatio],
+                    default=AspectRatio.r_16_9.value,
+                    tooltip="The aspect ratio of the output video",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=[model.value for model in Resolution],
+                    default=Resolution.r_1080p.value,
+                    tooltip="Supported values may vary by model & duration",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "movement_amplitude",
+                    options=[model.value for model in MovementAmplitude],
+                    default=MovementAmplitude.auto.value,
+                    tooltip="The movement amplitude of objects in the frame",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        prompt: str,
+        duration: int,
+        seed: int,
+        aspect_ratio: str,
+        resolution: str,
+        movement_amplitude: str,
+    ) -> comfy_io.NodeOutput:
+        if not prompt:
+            raise ValueError("The prompt field is required and cannot be empty.")
+        payload = TaskCreationRequest(
+            model_name=model,
+            prompt=prompt,
+            duration=duration,
+            seed=seed,
+            aspect_ratio=aspect_ratio,
+            resolution=resolution,
+            movement_amplitude=movement_amplitude,
+        )
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        results = await execute_task(VIDU_TEXT_TO_VIDEO, auth, payload, 320, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url))
+
+
+class ViduImageToVideoNode(comfy_io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="ViduImageToVideoNode",
+            display_name="Vidu Image To Video Generation",
+            category="api node/video/Vidu",
+            description="Generate video from image and optional prompt",
+            inputs=[
+                comfy_io.Combo.Input(
+                    "model",
+                    options=[model.value for model in VideoModelName],
+                    default=VideoModelName.vidu_q1.value,
+                    tooltip="Model name",
+                ),
+                comfy_io.Image.Input(
+                    "image",
+                    tooltip="An image to be used as the start frame of the generated video",
+                ),
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="A textual description for video generation",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=5,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=[model.value for model in Resolution],
+                    default=Resolution.r_1080p.value,
+                    tooltip="Supported values may vary by model & duration",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "movement_amplitude",
+                    options=[model.value for model in MovementAmplitude],
+                    default=MovementAmplitude.auto.value,
+                    tooltip="The movement amplitude of objects in the frame",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        image: torch.Tensor,
+        prompt: str,
+        duration: int,
+        seed: int,
+        resolution: str,
+        movement_amplitude: str,
+    ) -> comfy_io.NodeOutput:
+        if get_number_of_images(image) > 1:
+            raise ValueError("Only one input image is allowed.")
+        validate_image_aspect_ratio_range(image, (1, 4), (4, 1))
+        payload = TaskCreationRequest(
+            model_name=model,
+            prompt=prompt,
+            duration=duration,
+            seed=seed,
+            resolution=resolution,
+            movement_amplitude=movement_amplitude,
+        )
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        payload.images = await upload_images_to_comfyapi(
+            image,
+            max_images=1,
+            mime_type="image/png",
+            auth_kwargs=auth,
+        )
+        results = await execute_task(VIDU_IMAGE_TO_VIDEO, auth, payload, 120, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url))
+
+
+class ViduReferenceVideoNode(comfy_io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="ViduReferenceVideoNode",
+            display_name="Vidu Reference To Video Generation",
+            category="api node/video/Vidu",
+            description="Generate video from multiple images and prompt",
+            inputs=[
+                comfy_io.Combo.Input(
+                    "model",
+                    options=[model.value for model in VideoModelName],
+                    default=VideoModelName.vidu_q1.value,
+                    tooltip="Model name",
+                ),
+                comfy_io.Image.Input(
+                    "images",
+                    tooltip="Images to use as references to generate a video with consistent subjects (max 7 images).",
+                ),
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="A textual description for video generation",
+                ),
+                comfy_io.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=5,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=[model.value for model in AspectRatio],
+                    default=AspectRatio.r_16_9.value,
+                    tooltip="The aspect ratio of the output video",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=[model.value for model in Resolution],
+                    default=Resolution.r_1080p.value,
+                    tooltip="Supported values may vary by model & duration",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "movement_amplitude",
+                    options=[model.value for model in MovementAmplitude],
+                    default=MovementAmplitude.auto.value,
+                    tooltip="The movement amplitude of objects in the frame",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        images: torch.Tensor,
+        prompt: str,
+        duration: int,
+        seed: int,
+        aspect_ratio: str,
+        resolution: str,
+        movement_amplitude: str,
+    ) -> comfy_io.NodeOutput:
+        if not prompt:
+            raise ValueError("The prompt field is required and cannot be empty.")
+        a = get_number_of_images(images)
+        if a > 7:
+            raise ValueError("Too many images, maximum allowed is 7.")
+        for image in images:
+            validate_image_aspect_ratio_range(image, (1, 4), (4, 1))
+            validate_image_dimensions(image, min_width=128, min_height=128)
+        payload = TaskCreationRequest(
+            model_name=model,
+            prompt=prompt,
+            duration=duration,
+            seed=seed,
+            aspect_ratio=aspect_ratio,
+            resolution=resolution,
+            movement_amplitude=movement_amplitude,
+        )
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        payload.images = await upload_images_to_comfyapi(
+            images,
+            max_images=7,
+            mime_type="image/png",
+            auth_kwargs=auth,
+        )
+        results = await execute_task(VIDU_REFERENCE_VIDEO, auth, payload, 120, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url))
+
+
+class ViduStartEndToVideoNode(comfy_io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="ViduStartEndToVideoNode",
+            display_name="Vidu Start End To Video Generation",
+            category="api node/video/Vidu",
+            description="Generate a video from start and end frames and a prompt",
+            inputs=[
+                comfy_io.Combo.Input(
+                    "model",
+                    options=[model.value for model in VideoModelName],
+                    default=VideoModelName.vidu_q1.value,
+                    tooltip="Model name",
+                ),
+                comfy_io.Image.Input(
+                    "first_frame",
+                    tooltip="Start frame",
+                ),
+                comfy_io.Image.Input(
+                    "end_frame",
+                    tooltip="End frame",
+                ),
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="A textual description for video generation",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=5,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=[model.value for model in Resolution],
+                    default=Resolution.r_1080p.value,
+                    tooltip="Supported values may vary by model & duration",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "movement_amplitude",
+                    options=[model.value for model in MovementAmplitude],
+                    default=MovementAmplitude.auto.value,
+                    tooltip="The movement amplitude of objects in the frame",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        first_frame: torch.Tensor,
+        end_frame: torch.Tensor,
+        prompt: str,
+        duration: int,
+        seed: int,
+        resolution: str,
+        movement_amplitude: str,
+    ) -> comfy_io.NodeOutput:
+        validate_aspect_ratio_closeness(first_frame, end_frame, min_rel=0.8, max_rel=1.25, strict=False)
+        payload = TaskCreationRequest(
+            model_name=model,
+            prompt=prompt,
+            duration=duration,
+            seed=seed,
+            resolution=resolution,
+            movement_amplitude=movement_amplitude,
+        )
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        payload.images = [
+            (await upload_images_to_comfyapi(frame, max_images=1, mime_type="image/png", auth_kwargs=auth))[0]
+            for frame in (first_frame, end_frame)
+        ]
+        results = await execute_task(VIDU_START_END_VIDEO, auth, payload, 96, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url))
+
+
+class ViduExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[comfy_io.ComfyNode]]:
+        return [
+            ViduTextToVideoNode,
+            ViduImageToVideoNode,
+            ViduReferenceVideoNode,
+            ViduStartEndToVideoNode,
+        ]
+
+async def comfy_entrypoint() -> ViduExtension:
+    return ViduExtension()

comfy_api_nodes/util/validation_utils.py

@@ -53,6 +53,53 @@ def validate_image_aspect_ratio(
         )
 
 
+def validate_image_aspect_ratio_range(
+    image: torch.Tensor,
+    min_ratio: tuple[float, float],  # e.g. (1, 4)
+    max_ratio: tuple[float, float],  # e.g. (4, 1)
+    *,
+    strict: bool = True,             # True -> (min, max); False -> [min, max]
+) -> float:
+    a1, b1 = min_ratio
+    a2, b2 = max_ratio
+    if a1 <= 0 or b1 <= 0 or a2 <= 0 or b2 <= 0:
+        raise ValueError("Ratios must be positive, like (1, 4) or (4, 1).")
+    lo, hi = (a1 / b1), (a2 / b2)
+    if lo > hi:
+        lo, hi = hi, lo
+        a1, b1, a2, b2 = a2, b2, a1, b1  # swap only for error text
+    w, h = get_image_dimensions(image)
+    if w <= 0 or h <= 0:
+        raise ValueError(f"Invalid image dimensions: {w}x{h}")
+    ar = w / h
+    ok = (lo < ar < hi) if strict else (lo <= ar <= hi)
+    if not ok:
+        op = "<" if strict else "≤"
+        raise ValueError(f"Image aspect ratio {ar:.6g} is outside allowed range: {a1}:{b1} {op} ratio {op} {a2}:{b2}")
+    return ar
+
+
+def validate_aspect_ratio_closeness(
+    start_img,
+    end_img,
+    min_rel: float,
+    max_rel: float,
+    *,
+    strict: bool = False,   # True => exclusive, False => inclusive
+) -> None:
+    w1, h1 = get_image_dimensions(start_img)
+    w2, h2 = get_image_dimensions(end_img)
+    if min(w1, h1, w2, h2) <= 0:
+        raise ValueError("Invalid image dimensions")
+    ar1 = w1 / h1
+    ar2 = w2 / h2
+    # Normalize so it is symmetric (no need to check both ar1/ar2 and ar2/ar1)
+    closeness = max(ar1, ar2) / min(ar1, ar2)
+    limit = max(max_rel, 1.0 / min_rel)  # for 0.8..1.25 this is 1.25
+    if (closeness >= limit) if strict else (closeness > limit):
+        raise ValueError(f"Aspect ratios must be close: start/end={ar1/ar2:.4f}, allowed range {min_rel}–{max_rel}.")
+
+
 def validate_video_dimensions(
     video: VideoInput,
     min_width: Optional[int] = None,
@@ -98,3 +145,9 @@ def validate_video_duration(
         raise ValueError(
             f"Video duration must be at most {max_duration}s, got {duration}s"
         )
+
+
+def get_number_of_images(images):
+    if isinstance(images, torch.Tensor):
+        return images.shape[0] if images.ndim >= 4 else 1
+    return len(images)

comfy_extras/nodes_ace.py

@@ -1,49 +1,63 @@
 import torch
+from typing_extensions import override
+
 import comfy.model_management
 import node_helpers
+from comfy_api.latest import ComfyExtension, io
 
-class TextEncodeAceStepAudio:
+
+class TextEncodeAceStepAudio(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {
-            "clip": ("CLIP", ),
-            "tags": ("STRING", {"multiline": True, "dynamicPrompts": True}),
-            "lyrics": ("STRING", {"multiline": True, "dynamicPrompts": True}),
-            "lyrics_strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-            }}
-    RETURN_TYPES = ("CONDITIONING",)
-    FUNCTION = "encode"
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncodeAceStepAudio",
+            category="conditioning",
+            inputs=[
+                io.Clip.Input("clip"),
+                io.String.Input("tags", multiline=True, dynamic_prompts=True),
+                io.String.Input("lyrics", multiline=True, dynamic_prompts=True),
+                io.Float.Input("lyrics_strength", default=1.0, min=0.0, max=10.0, step=0.01),
+            ],
+            outputs=[io.Conditioning.Output()],
+        )
 
-    CATEGORY = "conditioning"
-
-    def encode(self, clip, tags, lyrics, lyrics_strength):
+    @classmethod
+    def execute(cls, clip, tags, lyrics, lyrics_strength) -> io.NodeOutput:
         tokens = clip.tokenize(tags, lyrics=lyrics)
         conditioning = clip.encode_from_tokens_scheduled(tokens)
         conditioning = node_helpers.conditioning_set_values(conditioning, {"lyrics_strength": lyrics_strength})
-        return (conditioning, )
+        return io.NodeOutput(conditioning)
 
 
-class EmptyAceStepLatentAudio:
-    def __init__(self):
-        self.device = comfy.model_management.intermediate_device()
+class EmptyAceStepLatentAudio(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="EmptyAceStepLatentAudio",
+            category="latent/audio",
+            inputs=[
+                io.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.1),
+                io.Int.Input(
+                    "batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch."
+                ),
+            ],
+            outputs=[io.Latent.Output()],
+        )
 
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"seconds": ("FLOAT", {"default": 120.0, "min": 1.0, "max": 1000.0, "step": 0.1}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096, "tooltip": "The number of latent images in the batch."}),
-                             }}
-    RETURN_TYPES = ("LATENT",)
-    FUNCTION = "generate"
-
-    CATEGORY = "latent/audio"
-
-    def generate(self, seconds, batch_size):
+    def execute(cls, seconds, batch_size) -> io.NodeOutput:
         length = int(seconds * 44100 / 512 / 8)
-        latent = torch.zeros([batch_size, 8, 16, length], device=self.device)
-        return ({"samples": latent, "type": "audio"}, )
+        latent = torch.zeros([batch_size, 8, 16, length], device=comfy.model_management.intermediate_device())
+        return io.NodeOutput({"samples": latent, "type": "audio"})
 
 
-NODE_CLASS_MAPPINGS = {
-    "TextEncodeAceStepAudio": TextEncodeAceStepAudio,
-    "EmptyAceStepLatentAudio": EmptyAceStepLatentAudio,
-}
+class AceExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            TextEncodeAceStepAudio,
+            EmptyAceStepLatentAudio,
+        ]
+
+async def comfy_entrypoint() -> AceExtension:
+    return AceExtension()

comfy_extras/nodes_advanced_samplers.py

@@ -1,8 +1,13 @@
+import numpy as np
+import torch
+from tqdm.auto import trange
+from typing_extensions import override
+
+import comfy.model_patcher
 import comfy.samplers
 import comfy.utils
-import torch
-import numpy as np
-from tqdm.auto import trange
+from comfy.k_diffusion.sampling import to_d
+from comfy_api.latest import ComfyExtension, io
 
 
 @torch.no_grad()
@@ -33,30 +38,29 @@ def sample_lcm_upscale(model, x, sigmas, extra_args=None, callback=None, disable
     return x
 
 
-class SamplerLCMUpscale:
-    upscale_methods = ["bislerp", "nearest-exact", "bilinear", "area", "bicubic"]
+class SamplerLCMUpscale(io.ComfyNode):
+    UPSCALE_METHODS = ["bislerp", "nearest-exact", "bilinear", "area", "bicubic"]
 
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required":
-                    {"scale_ratio": ("FLOAT", {"default": 1.0, "min": 0.1, "max": 20.0, "step": 0.01}),
-                     "scale_steps": ("INT", {"default": -1, "min": -1, "max": 1000, "step": 1}),
-                     "upscale_method": (s.upscale_methods,),
-                      }
-               }
-    RETURN_TYPES = ("SAMPLER",)
-    CATEGORY = "sampling/custom_sampling/samplers"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="SamplerLCMUpscale",
+            category="sampling/custom_sampling/samplers",
+            inputs=[
+                io.Float.Input("scale_ratio", default=1.0, min=0.1, max=20.0, step=0.01),
+                io.Int.Input("scale_steps", default=-1, min=-1, max=1000, step=1),
+                io.Combo.Input("upscale_method", options=cls.UPSCALE_METHODS),
+            ],
+            outputs=[io.Sampler.Output()],
+        )
 
-    FUNCTION = "get_sampler"
-
-    def get_sampler(self, scale_ratio, scale_steps, upscale_method):
+    @classmethod
+    def execute(cls, scale_ratio, scale_steps, upscale_method) -> io.NodeOutput:
         if scale_steps < 0:
             scale_steps = None
         sampler = comfy.samplers.KSAMPLER(sample_lcm_upscale, extra_options={"total_upscale": scale_ratio, "upscale_steps": scale_steps, "upscale_method": upscale_method})
-        return (sampler, )
+        return io.NodeOutput(sampler)
 
-from comfy.k_diffusion.sampling import to_d
-import comfy.model_patcher
 
 @torch.no_grad()
 def sample_euler_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
@@ -82,30 +86,36 @@ def sample_euler_pp(model, x, sigmas, extra_args=None, callback=None, disable=No
     return x
 
 
-class SamplerEulerCFGpp:
+class SamplerEulerCFGpp(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required":
-                    {"version": (["regular", "alternative"],),}
-               }
-    RETURN_TYPES = ("SAMPLER",)
-    # CATEGORY = "sampling/custom_sampling/samplers"
-    CATEGORY = "_for_testing"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="SamplerEulerCFGpp",
+            display_name="SamplerEulerCFG++",
+            category="_for_testing",  # "sampling/custom_sampling/samplers"
+            inputs=[
+                io.Combo.Input("version", options=["regular", "alternative"]),
+            ],
+            outputs=[io.Sampler.Output()],
+            is_experimental=True,
+        )
 
-    FUNCTION = "get_sampler"
-
-    def get_sampler(self, version):
+    @classmethod
+    def execute(cls, version) -> io.NodeOutput:
         if version == "alternative":
             sampler = comfy.samplers.KSAMPLER(sample_euler_pp)
         else:
             sampler = comfy.samplers.ksampler("euler_cfg_pp")
-        return (sampler, )
+        return io.NodeOutput(sampler)
 
-NODE_CLASS_MAPPINGS = {
-    "SamplerLCMUpscale": SamplerLCMUpscale,
-    "SamplerEulerCFGpp": SamplerEulerCFGpp,
-}
 
-NODE_DISPLAY_NAME_MAPPINGS = {
-    "SamplerEulerCFGpp": "SamplerEulerCFG++",
-}
+class AdvancedSamplersExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            SamplerLCMUpscale,
+            SamplerEulerCFGpp,
+        ]
+
+async def comfy_entrypoint() -> AdvancedSamplersExtension:
+    return AdvancedSamplersExtension()

comfy_extras/nodes_apg.py

@@ -1,4 +1,8 @@
 import torch
+from typing_extensions import override
+
+from comfy_api.latest import ComfyExtension, io
+
 
 def project(v0, v1):
     v1 = torch.nn.functional.normalize(v1, dim=[-1, -2, -3])
@@ -6,22 +10,45 @@ def project(v0, v1):
     v0_orthogonal = v0 - v0_parallel
     return v0_parallel, v0_orthogonal
 
-class APG:
+class APG(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "model": ("MODEL",),
-                "eta": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01, "tooltip": "Controls the scale of the parallel guidance vector. Default CFG behavior at a setting of 1."}),
-                "norm_threshold": ("FLOAT", {"default": 5.0, "min": 0.0, "max": 50.0, "step": 0.1, "tooltip": "Normalize guidance vector to this value, normalization disable at a setting of 0."}),
-                "momentum": ("FLOAT", {"default": 0.0, "min": -5.0, "max": 1.0, "step": 0.01, "tooltip":"Controls a running average of guidance during diffusion, disabled at a setting of 0."}),
-            }
-        }
-    RETURN_TYPES = ("MODEL",)
-    FUNCTION = "patch"
-    CATEGORY = "sampling/custom_sampling"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="APG",
+            display_name="Adaptive Projected Guidance",
+            category="sampling/custom_sampling",
+            inputs=[
+                io.Model.Input("model"),
+                io.Float.Input(
+                    "eta",
+                    default=1.0,
+                    min=-10.0,
+                    max=10.0,
+                    step=0.01,
+                    tooltip="Controls the scale of the parallel guidance vector. Default CFG behavior at a setting of 1.",
+                ),
+                io.Float.Input(
+                    "norm_threshold",
+                    default=5.0,
+                    min=0.0,
+                    max=50.0,
+                    step=0.1,
+                    tooltip="Normalize guidance vector to this value, normalization disable at a setting of 0.",
+                ),
+                io.Float.Input(
+                    "momentum",
+                    default=0.0,
+                    min=-5.0,
+                    max=1.0,
+                    step=0.01,
+                    tooltip="Controls a running average of guidance during diffusion, disabled at a setting of 0.",
+                ),
+            ],
+            outputs=[io.Model.Output()],
+        )
 
-    def patch(self, model, eta, norm_threshold, momentum):
+    @classmethod
+    def execute(cls, model, eta, norm_threshold, momentum) -> io.NodeOutput:
         running_avg = 0
         prev_sigma = None
 
@@ -65,12 +92,15 @@ class APG:
 
         m = model.clone()
         m.set_model_sampler_pre_cfg_function(pre_cfg_function)
-        return (m,)
+        return io.NodeOutput(m)
 
-NODE_CLASS_MAPPINGS = {
-    "APG": APG,
-}
 
-NODE_DISPLAY_NAME_MAPPINGS = {
-    "APG": "Adaptive Projected Guidance",
-}
+class ApgExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            APG,
+        ]
+
+async def comfy_entrypoint() -> ApgExtension:
+    return ApgExtension()

comfy_extras/nodes_attention_multiply.py

@@ -1,3 +1,7 @@
+from typing_extensions import override
+
+from comfy_api.latest import ComfyExtension, io
+
 
 def attention_multiply(attn, model, q, k, v, out):
     m = model.clone()
@@ -16,57 +20,71 @@ def attention_multiply(attn, model, q, k, v, out):
     return m
 
 
-class UNetSelfAttentionMultiply:
+class UNetSelfAttentionMultiply(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "model": ("MODEL",),
-                              "q": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "k": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "v": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "out": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              }}
-    RETURN_TYPES = ("MODEL",)
-    FUNCTION = "patch"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="UNetSelfAttentionMultiply",
+            category="_for_testing/attention_experiments",
+            inputs=[
+                io.Model.Input("model"),
+                io.Float.Input("q", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("k", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("v", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("out", default=1.0, min=0.0, max=10.0, step=0.01),
+            ],
+            outputs=[io.Model.Output()],
+            is_experimental=True,
+        )
 
-    CATEGORY = "_for_testing/attention_experiments"
-
-    def patch(self, model, q, k, v, out):
+    @classmethod
+    def execute(cls, model, q, k, v, out) -> io.NodeOutput:
         m = attention_multiply("attn1", model, q, k, v, out)
-        return (m, )
+        return io.NodeOutput(m)
 
-class UNetCrossAttentionMultiply:
+
+class UNetCrossAttentionMultiply(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "model": ("MODEL",),
-                              "q": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "k": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "v": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "out": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              }}
-    RETURN_TYPES = ("MODEL",)
-    FUNCTION = "patch"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="UNetCrossAttentionMultiply",
+            category="_for_testing/attention_experiments",
+            inputs=[
+                io.Model.Input("model"),
+                io.Float.Input("q", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("k", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("v", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("out", default=1.0, min=0.0, max=10.0, step=0.01),
+            ],
+            outputs=[io.Model.Output()],
+            is_experimental=True,
+        )
 
-    CATEGORY = "_for_testing/attention_experiments"
-
-    def patch(self, model, q, k, v, out):
+    @classmethod
+    def execute(cls, model, q, k, v, out) -> io.NodeOutput:
         m = attention_multiply("attn2", model, q, k, v, out)
-        return (m, )
+        return io.NodeOutput(m)
 
-class CLIPAttentionMultiply:
+
+class CLIPAttentionMultiply(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "clip": ("CLIP",),
-                              "q": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "k": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "v": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "out": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              }}
-    RETURN_TYPES = ("CLIP",)
-    FUNCTION = "patch"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="CLIPAttentionMultiply",
+            category="_for_testing/attention_experiments",
+            inputs=[
+                io.Clip.Input("clip"),
+                io.Float.Input("q", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("k", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("v", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("out", default=1.0, min=0.0, max=10.0, step=0.01),
+            ],
+            outputs=[io.Clip.Output()],
+            is_experimental=True,
+        )
 
-    CATEGORY = "_for_testing/attention_experiments"
-
-    def patch(self, clip, q, k, v, out):
+    @classmethod
+    def execute(cls, clip, q, k, v, out) -> io.NodeOutput:
         m = clip.clone()
         sd = m.patcher.model_state_dict()
 
@@ -79,23 +97,28 @@ class CLIPAttentionMultiply:
                 m.add_patches({key: (None,)}, 0.0, v)
             if key.endswith("self_attn.out_proj.weight") or key.endswith("self_attn.out_proj.bias"):
                 m.add_patches({key: (None,)}, 0.0, out)
-        return (m, )
+        return io.NodeOutput(m)
 
-class UNetTemporalAttentionMultiply:
+
+class UNetTemporalAttentionMultiply(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "model": ("MODEL",),
-                              "self_structural": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "self_temporal": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "cross_structural": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "cross_temporal": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              }}
-    RETURN_TYPES = ("MODEL",)
-    FUNCTION = "patch"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="UNetTemporalAttentionMultiply",
+            category="_for_testing/attention_experiments",
+            inputs=[
+                io.Model.Input("model"),
+                io.Float.Input("self_structural", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("self_temporal", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("cross_structural", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("cross_temporal", default=1.0, min=0.0, max=10.0, step=0.01),
+            ],
+            outputs=[io.Model.Output()],
+            is_experimental=True,
+        )
 
-    CATEGORY = "_for_testing/attention_experiments"
-
-    def patch(self, model, self_structural, self_temporal, cross_structural, cross_temporal):
+    @classmethod
+    def execute(cls, model, self_structural, self_temporal, cross_structural, cross_temporal) -> io.NodeOutput:
         m = model.clone()
         sd = model.model_state_dict()
 
@@ -110,11 +133,18 @@ class UNetTemporalAttentionMultiply:
                     m.add_patches({k: (None,)}, 0.0, cross_temporal)
                 else:
                     m.add_patches({k: (None,)}, 0.0, cross_structural)
-        return (m, )
+        return io.NodeOutput(m)
 
-NODE_CLASS_MAPPINGS = {
-    "UNetSelfAttentionMultiply": UNetSelfAttentionMultiply,
-    "UNetCrossAttentionMultiply": UNetCrossAttentionMultiply,
-    "CLIPAttentionMultiply": CLIPAttentionMultiply,
-    "UNetTemporalAttentionMultiply": UNetTemporalAttentionMultiply,
-}
+
+class AttentionMultiplyExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            UNetSelfAttentionMultiply,
+            UNetCrossAttentionMultiply,
+            CLIPAttentionMultiply,
+            UNetTemporalAttentionMultiply,
+        ]
+
+async def comfy_entrypoint() -> AttentionMultiplyExtension:
+    return AttentionMultiplyExtension()

comfy_extras/nodes_audio.py

@@ -346,6 +346,24 @@ class LoadAudio:
             return "Invalid audio file: {}".format(audio)
         return True
 
+class RecordAudio:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"audio": ("AUDIO_RECORD", {})}}
+
+    CATEGORY = "audio"
+
+    RETURN_TYPES = ("AUDIO", )
+    FUNCTION = "load"
+
+    def load(self, audio):
+        audio_path = folder_paths.get_annotated_filepath(audio)
+
+        waveform, sample_rate = torchaudio.load(audio_path)
+        audio = {"waveform": waveform.unsqueeze(0), "sample_rate": sample_rate}
+        return (audio, )
+
+
 NODE_CLASS_MAPPINGS = {
     "EmptyLatentAudio": EmptyLatentAudio,
     "VAEEncodeAudio": VAEEncodeAudio,
@@ -356,6 +374,7 @@ NODE_CLASS_MAPPINGS = {
     "LoadAudio": LoadAudio,
     "PreviewAudio": PreviewAudio,
     "ConditioningStableAudio": ConditioningStableAudio,
+    "RecordAudio": RecordAudio,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
@@ -367,4 +386,5 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "SaveAudio": "Save Audio (FLAC)",
     "SaveAudioMP3": "Save Audio (MP3)",
     "SaveAudioOpus": "Save Audio (Opus)",
+    "RecordAudio": "Record Audio",
 }

comfy_extras/nodes_audio_encoder.py

@@ -0,0 +1,44 @@
+import folder_paths
+import comfy.audio_encoders.audio_encoders
+import comfy.utils
+
+
+class AudioEncoderLoader:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "audio_encoder_name": (folder_paths.get_filename_list("audio_encoders"), ),
+                             }}
+    RETURN_TYPES = ("AUDIO_ENCODER",)
+    FUNCTION = "load_model"
+
+    CATEGORY = "loaders"
+
+    def load_model(self, audio_encoder_name):
+        audio_encoder_name = folder_paths.get_full_path_or_raise("audio_encoders", audio_encoder_name)
+        sd = comfy.utils.load_torch_file(audio_encoder_name, safe_load=True)
+        audio_encoder = comfy.audio_encoders.audio_encoders.load_audio_encoder_from_sd(sd)
+        if audio_encoder is None:
+            raise RuntimeError("ERROR: audio encoder file is invalid and does not contain a valid model.")
+        return (audio_encoder,)
+
+
+class AudioEncoderEncode:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "audio_encoder": ("AUDIO_ENCODER",),
+                              "audio": ("AUDIO",),
+                             }}
+    RETURN_TYPES = ("AUDIO_ENCODER_OUTPUT",)
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning"
+
+    def encode(self, audio_encoder, audio):
+        output = audio_encoder.encode_audio(audio["waveform"], audio["sample_rate"])
+        return (output,)
+
+
+NODE_CLASS_MAPPINGS = {
+    "AudioEncoderLoader": AudioEncoderLoader,
+    "AudioEncoderEncode": AudioEncoderEncode,
+}

comfy_extras/nodes_context_windows.py

@@ -0,0 +1,89 @@
+from __future__ import annotations
+from comfy_api.latest import ComfyExtension, io
+import comfy.context_windows
+import nodes
+
+
+class ContextWindowsManualNode(io.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="ContextWindowsManual",
+            display_name="Context Windows (Manual)",
+            category="context",
+            description="Manually set context windows.",
+            inputs=[
+                io.Model.Input("model", tooltip="The model to apply context windows to during sampling."),
+                io.Int.Input("context_length", min=1, default=16, tooltip="The length of the context window."),
+                io.Int.Input("context_overlap", min=0, default=4, tooltip="The overlap of the context window."),
+                io.Combo.Input("context_schedule", options=[
+                    comfy.context_windows.ContextSchedules.STATIC_STANDARD,
+                    comfy.context_windows.ContextSchedules.UNIFORM_STANDARD,
+                    comfy.context_windows.ContextSchedules.UNIFORM_LOOPED,
+                    comfy.context_windows.ContextSchedules.BATCHED,
+                    ], tooltip="The stride of the context window."),
+                io.Int.Input("context_stride", min=1, default=1, tooltip="The stride of the context window; only applicable to uniform schedules."),
+                io.Boolean.Input("closed_loop", default=False, tooltip="Whether to close the context window loop; only applicable to looped schedules."),
+                io.Combo.Input("fuse_method", options=comfy.context_windows.ContextFuseMethods.LIST_STATIC, default=comfy.context_windows.ContextFuseMethods.PYRAMID, tooltip="The method to use to fuse the context windows."),
+                io.Int.Input("dim", min=0, max=5, default=0, tooltip="The dimension to apply the context windows to."),
+            ],
+            outputs=[
+                io.Model.Output(tooltip="The model with context windows applied during sampling."),
+            ],
+            is_experimental=True,
+        )
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str, dim: int) -> io.Model:
+        model = model.clone()
+        model.model_options["context_handler"] = comfy.context_windows.IndexListContextHandler(
+            context_schedule=comfy.context_windows.get_matching_context_schedule(context_schedule),
+            fuse_method=comfy.context_windows.get_matching_fuse_method(fuse_method),
+            context_length=context_length,
+            context_overlap=context_overlap,
+            context_stride=context_stride,
+            closed_loop=closed_loop,
+            dim=dim)
+        # make memory usage calculation only take into account the context window latents
+        comfy.context_windows.create_prepare_sampling_wrapper(model)
+        return io.NodeOutput(model)
+
+class WanContextWindowsManualNode(ContextWindowsManualNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        schema = super().define_schema()
+        schema.node_id = "WanContextWindowsManual"
+        schema.display_name = "WAN Context Windows (Manual)"
+        schema.description = "Manually set context windows for WAN-like models (dim=2)."
+        schema.inputs = [
+            io.Model.Input("model", tooltip="The model to apply context windows to during sampling."),
+                io.Int.Input("context_length", min=1, max=nodes.MAX_RESOLUTION, step=4, default=81, tooltip="The length of the context window."),
+                io.Int.Input("context_overlap", min=0, default=30, tooltip="The overlap of the context window."),
+                io.Combo.Input("context_schedule", options=[
+                    comfy.context_windows.ContextSchedules.STATIC_STANDARD,
+                    comfy.context_windows.ContextSchedules.UNIFORM_STANDARD,
+                    comfy.context_windows.ContextSchedules.UNIFORM_LOOPED,
+                    comfy.context_windows.ContextSchedules.BATCHED,
+                    ], tooltip="The stride of the context window."),
+                io.Int.Input("context_stride", min=1, default=1, tooltip="The stride of the context window; only applicable to uniform schedules."),
+                io.Boolean.Input("closed_loop", default=False, tooltip="Whether to close the context window loop; only applicable to looped schedules."),
+                io.Combo.Input("fuse_method", options=comfy.context_windows.ContextFuseMethods.LIST_STATIC, default=comfy.context_windows.ContextFuseMethods.PYRAMID, tooltip="The method to use to fuse the context windows."),
+        ]
+        return schema
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str) -> io.Model:
+        context_length = max(((context_length - 1) // 4) + 1, 1)  # at least length 1
+        context_overlap = max(((context_overlap - 1) // 4) + 1, 0)  # at least overlap 0
+        return super().execute(model, context_length, context_overlap, context_schedule, context_stride, closed_loop, fuse_method, dim=2)
+
+
+class ContextWindowsExtension(ComfyExtension):
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            ContextWindowsManualNode,
+            WanContextWindowsManualNode,
+        ]
+
+def comfy_entrypoint():
+    return ContextWindowsExtension()

comfy_extras/nodes_easycache.py

@@ -0,0 +1,493 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING, Union
+from comfy_api.latest import io, ComfyExtension
+import comfy.patcher_extension
+import logging
+import torch
+import comfy.model_patcher
+if TYPE_CHECKING:
+    from uuid import UUID
+
+
+def easycache_forward_wrapper(executor, *args, **kwargs):
+    # get values from args
+    x: torch.Tensor = args[0]
+    transformer_options: dict[str] = args[-1]
+    if not isinstance(transformer_options, dict):
+        transformer_options = kwargs.get("transformer_options")
+        if not transformer_options:
+            transformer_options = args[-2]
+    easycache: EasyCacheHolder = transformer_options["easycache"]
+    sigmas = transformer_options["sigmas"]
+    uuids = transformer_options["uuids"]
+    if sigmas is not None and easycache.is_past_end_timestep(sigmas):
+        return executor(*args, **kwargs)
+    # prepare next x_prev
+    has_first_cond_uuid = easycache.has_first_cond_uuid(uuids)
+    next_x_prev = x
+    input_change = None
+    do_easycache = easycache.should_do_easycache(sigmas)
+    if do_easycache:
+        easycache.check_metadata(x)
+        # if first cond marked this step for skipping, skip it and use appropriate cached values
+        if easycache.skip_current_step:
+            if easycache.verbose:
+                logging.info(f"EasyCache [verbose] - was marked to skip this step by {easycache.first_cond_uuid}. Present uuids: {uuids}")
+            return easycache.apply_cache_diff(x, uuids)
+        if easycache.initial_step:
+            easycache.first_cond_uuid = uuids[0]
+            has_first_cond_uuid = easycache.has_first_cond_uuid(uuids)
+            easycache.initial_step = False
+        if has_first_cond_uuid:
+            if easycache.has_x_prev_subsampled():
+                input_change = (easycache.subsample(x, uuids, clone=False) - easycache.x_prev_subsampled).flatten().abs().mean()
+            if easycache.has_output_prev_norm() and easycache.has_relative_transformation_rate():
+                approx_output_change_rate = (easycache.relative_transformation_rate * input_change) / easycache.output_prev_norm
+                easycache.cumulative_change_rate += approx_output_change_rate
+                if easycache.cumulative_change_rate < easycache.reuse_threshold:
+                    if easycache.verbose:
+                        logging.info(f"EasyCache [verbose] - skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
+                    # other conds should also skip this step, and instead use their cached values
+                    easycache.skip_current_step = True
+                    return easycache.apply_cache_diff(x, uuids)
+                else:
+                    if easycache.verbose:
+                        logging.info(f"EasyCache [verbose] - NOT skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
+                    easycache.cumulative_change_rate = 0.0
+
+    output: torch.Tensor = executor(*args, **kwargs)
+    if has_first_cond_uuid and easycache.has_output_prev_norm():
+        output_change = (easycache.subsample(output, uuids, clone=False) - easycache.output_prev_subsampled).flatten().abs().mean()
+        if easycache.verbose:
+            output_change_rate = output_change / easycache.output_prev_norm
+            easycache.output_change_rates.append(output_change_rate.item())
+        if easycache.has_relative_transformation_rate():
+            approx_output_change_rate = (easycache.relative_transformation_rate * input_change) / easycache.output_prev_norm
+            easycache.approx_output_change_rates.append(approx_output_change_rate.item())
+            if easycache.verbose:
+                logging.info(f"EasyCache [verbose] - approx_output_change_rate: {approx_output_change_rate}")
+        if input_change is not None:
+            easycache.relative_transformation_rate = output_change / input_change
+        if easycache.verbose:
+            logging.info(f"EasyCache [verbose] - output_change_rate: {output_change_rate}")
+    # TODO: allow cache_diff to be offloaded
+    easycache.update_cache_diff(output, next_x_prev, uuids)
+    if has_first_cond_uuid:
+        easycache.x_prev_subsampled = easycache.subsample(next_x_prev, uuids)
+        easycache.output_prev_subsampled = easycache.subsample(output, uuids)
+        easycache.output_prev_norm = output.flatten().abs().mean()
+        if easycache.verbose:
+            logging.info(f"EasyCache [verbose] - x_prev_subsampled: {easycache.x_prev_subsampled.shape}")
+    return output
+
+def lazycache_predict_noise_wrapper(executor, *args, **kwargs):
+    # get values from args
+    x: torch.Tensor = args[0]
+    timestep: float = args[1]
+    model_options: dict[str] = args[2]
+    easycache: LazyCacheHolder = model_options["transformer_options"]["easycache"]
+    if easycache.is_past_end_timestep(timestep):
+        return executor(*args, **kwargs)
+    # prepare next x_prev
+    next_x_prev = x
+    input_change = None
+    do_easycache = easycache.should_do_easycache(timestep)
+    if do_easycache:
+        easycache.check_metadata(x)
+        if easycache.has_x_prev_subsampled():
+            if easycache.has_x_prev_subsampled():
+                input_change = (easycache.subsample(x, clone=False) - easycache.x_prev_subsampled).flatten().abs().mean()
+            if easycache.has_output_prev_norm() and easycache.has_relative_transformation_rate():
+                approx_output_change_rate = (easycache.relative_transformation_rate * input_change) / easycache.output_prev_norm
+                easycache.cumulative_change_rate += approx_output_change_rate
+                if easycache.cumulative_change_rate < easycache.reuse_threshold:
+                    if easycache.verbose:
+                        logging.info(f"LazyCache [verbose] - skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
+                    # other conds should also skip this step, and instead use their cached values
+                    easycache.skip_current_step = True
+                    return easycache.apply_cache_diff(x)
+                else:
+                    if easycache.verbose:
+                        logging.info(f"LazyCache [verbose] - NOT skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
+                    easycache.cumulative_change_rate = 0.0
+    output: torch.Tensor = executor(*args, **kwargs)
+    if easycache.has_output_prev_norm():
+        output_change = (easycache.subsample(output, clone=False) - easycache.output_prev_subsampled).flatten().abs().mean()
+        if easycache.verbose:
+            output_change_rate = output_change / easycache.output_prev_norm
+            easycache.output_change_rates.append(output_change_rate.item())
+        if easycache.has_relative_transformation_rate():
+            approx_output_change_rate = (easycache.relative_transformation_rate * input_change) / easycache.output_prev_norm
+            easycache.approx_output_change_rates.append(approx_output_change_rate.item())
+            if easycache.verbose:
+                logging.info(f"LazyCache [verbose] - approx_output_change_rate: {approx_output_change_rate}")
+        if input_change is not None:
+            easycache.relative_transformation_rate = output_change / input_change
+        if easycache.verbose:
+            logging.info(f"LazyCache [verbose] - output_change_rate: {output_change_rate}")
+    # TODO: allow cache_diff to be offloaded
+    easycache.update_cache_diff(output, next_x_prev)
+    easycache.x_prev_subsampled = easycache.subsample(next_x_prev)
+    easycache.output_prev_subsampled = easycache.subsample(output)
+    easycache.output_prev_norm = output.flatten().abs().mean()
+    if easycache.verbose:
+        logging.info(f"LazyCache [verbose] - x_prev_subsampled: {easycache.x_prev_subsampled.shape}")
+    return output
+
+def easycache_calc_cond_batch_wrapper(executor, *args, **kwargs):
+    model_options = args[-1]
+    easycache: EasyCacheHolder = model_options["transformer_options"]["easycache"]
+    easycache.skip_current_step = False
+    # TODO: check if first_cond_uuid is active at this timestep; otherwise, EasyCache needs to be partially reset
+    return executor(*args, **kwargs)
+
+def easycache_sample_wrapper(executor, *args, **kwargs):
+    """
+    This OUTER_SAMPLE wrapper makes sure easycache is prepped for current run, and all memory usage is cleared at the end.
+    """
+    try:
+        guider = executor.class_obj
+        orig_model_options = guider.model_options
+        guider.model_options = comfy.model_patcher.create_model_options_clone(orig_model_options)
+        # clone and prepare timesteps
+        guider.model_options["transformer_options"]["easycache"] = guider.model_options["transformer_options"]["easycache"].clone().prepare_timesteps(guider.model_patcher.model.model_sampling)
+        easycache: Union[EasyCacheHolder, LazyCacheHolder] = guider.model_options['transformer_options']['easycache']
+        logging.info(f"{easycache.name} enabled - threshold: {easycache.reuse_threshold}, start_percent: {easycache.start_percent}, end_percent: {easycache.end_percent}")
+        return executor(*args, **kwargs)
+    finally:
+        easycache = guider.model_options['transformer_options']['easycache']
+        output_change_rates = easycache.output_change_rates
+        approx_output_change_rates = easycache.approx_output_change_rates
+        if easycache.verbose:
+            logging.info(f"{easycache.name} [verbose] - output_change_rates {len(output_change_rates)}: {output_change_rates}")
+            logging.info(f"{easycache.name} [verbose] - approx_output_change_rates {len(approx_output_change_rates)}: {approx_output_change_rates}")
+        total_steps = len(args[3])-1
+        logging.info(f"{easycache.name} - skipped {easycache.total_steps_skipped}/{total_steps} steps ({total_steps/(total_steps-easycache.total_steps_skipped):.2f}x speedup).")
+        easycache.reset()
+        guider.model_options = orig_model_options
+
+
+class EasyCacheHolder:
+    def __init__(self, reuse_threshold: float, start_percent: float, end_percent: float, subsample_factor: int, offload_cache_diff: bool, verbose: bool=False):
+        self.name = "EasyCache"
+        self.reuse_threshold = reuse_threshold
+        self.start_percent = start_percent
+        self.end_percent = end_percent
+        self.subsample_factor = subsample_factor
+        self.offload_cache_diff = offload_cache_diff
+        self.verbose = verbose
+        # timestep values
+        self.start_t = 0.0
+        self.end_t = 0.0
+        # control values
+        self.relative_transformation_rate: float = None
+        self.cumulative_change_rate = 0.0
+        self.initial_step = True
+        self.skip_current_step = False
+        # cache values
+        self.first_cond_uuid = None
+        self.x_prev_subsampled: torch.Tensor = None
+        self.output_prev_subsampled: torch.Tensor = None
+        self.output_prev_norm: torch.Tensor = None
+        self.uuid_cache_diffs: dict[UUID, torch.Tensor] = {}
+        self.output_change_rates = []
+        self.approx_output_change_rates = []
+        self.total_steps_skipped = 0
+        # how to deal with mismatched dims
+        self.allow_mismatch = True
+        self.cut_from_start = True
+        self.state_metadata = None
+
+    def is_past_end_timestep(self, timestep: float) -> bool:
+        return not (timestep[0] > self.end_t).item()
+
+    def should_do_easycache(self, timestep: float) -> bool:
+        return (timestep[0] <= self.start_t).item()
+
+    def has_x_prev_subsampled(self) -> bool:
+        return self.x_prev_subsampled is not None
+
+    def has_output_prev_subsampled(self) -> bool:
+        return self.output_prev_subsampled is not None
+
+    def has_output_prev_norm(self) -> bool:
+        return self.output_prev_norm is not None
+
+    def has_relative_transformation_rate(self) -> bool:
+        return self.relative_transformation_rate is not None
+
+    def prepare_timesteps(self, model_sampling):
+        self.start_t = model_sampling.percent_to_sigma(self.start_percent)
+        self.end_t = model_sampling.percent_to_sigma(self.end_percent)
+        return self
+
+    def subsample(self, x: torch.Tensor, uuids: list[UUID], clone: bool = True) -> torch.Tensor:
+        batch_offset = x.shape[0] // len(uuids)
+        uuid_idx = uuids.index(self.first_cond_uuid)
+        if self.subsample_factor > 1:
+            to_return = x[uuid_idx*batch_offset:(uuid_idx+1)*batch_offset, ..., ::self.subsample_factor, ::self.subsample_factor]
+            if clone:
+                return to_return.clone()
+            return to_return
+        to_return = x[uuid_idx*batch_offset:(uuid_idx+1)*batch_offset, ...]
+        if clone:
+            return to_return.clone()
+        return to_return
+
+    def apply_cache_diff(self, x: torch.Tensor, uuids: list[UUID]):
+        if self.first_cond_uuid in uuids:
+            self.total_steps_skipped += 1
+        batch_offset = x.shape[0] // len(uuids)
+        for i, uuid in enumerate(uuids):
+            # if cached dims don't match x dims, cut off excess and hope for the best (cosmos world2video)
+            if x.shape[1:] != self.uuid_cache_diffs[uuid].shape[1:]:
+                if not self.allow_mismatch:
+                    raise ValueError(f"Cached dims {self.uuid_cache_diffs[uuid].shape} don't match x dims {x.shape} - this is no good")
+                slicing = []
+                skip_this_dim = True
+                for dim_u, dim_x in zip(self.uuid_cache_diffs[uuid].shape, x.shape):
+                    if skip_this_dim:
+                        skip_this_dim = False
+                        continue
+                    if dim_u != dim_x:
+                        if self.cut_from_start:
+                            slicing.append(slice(dim_x-dim_u, None))
+                        else:
+                            slicing.append(slice(None, dim_u))
+                    else:
+                        slicing.append(slice(None))
+                slicing = [slice(i*batch_offset,(i+1)*batch_offset)] + slicing
+                x = x[slicing]
+            x += self.uuid_cache_diffs[uuid].to(x.device)
+        return x
+
+    def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID]):
+        # if output dims don't match x dims, cut off excess and hope for the best (cosmos world2video)
+        if output.shape[1:] != x.shape[1:]:
+            if not self.allow_mismatch:
+                raise ValueError(f"Output dims {output.shape} don't match x dims {x.shape} - this is no good")
+            slicing = []
+            skip_dim = True
+            for dim_o, dim_x in zip(output.shape, x.shape):
+                if not skip_dim and dim_o != dim_x:
+                    if self.cut_from_start:
+                        slicing.append(slice(dim_x-dim_o, None))
+                    else:
+                        slicing.append(slice(None, dim_o))
+                else:
+                    slicing.append(slice(None))
+                skip_dim = False
+            x = x[slicing]
+        diff = output - x
+        batch_offset = diff.shape[0] // len(uuids)
+        for i, uuid in enumerate(uuids):
+            self.uuid_cache_diffs[uuid] = diff[i*batch_offset:(i+1)*batch_offset, ...]
+
+    def has_first_cond_uuid(self, uuids: list[UUID]) -> bool:
+        return self.first_cond_uuid in uuids
+
+    def check_metadata(self, x: torch.Tensor) -> bool:
+        metadata = (x.device, x.dtype, x.shape[1:])
+        if self.state_metadata is None:
+            self.state_metadata = metadata
+            return True
+        if metadata == self.state_metadata:
+            return True
+        logging.warn(f"{self.name} - Tensor shape, dtype or device changed, resetting state")
+        self.reset()
+        return False
+
+    def reset(self):
+        self.relative_transformation_rate = 0.0
+        self.cumulative_change_rate = 0.0
+        self.initial_step = True
+        self.skip_current_step = False
+        self.output_change_rates = []
+        self.first_cond_uuid = None
+        del self.x_prev_subsampled
+        self.x_prev_subsampled = None
+        del self.output_prev_subsampled
+        self.output_prev_subsampled = None
+        del self.output_prev_norm
+        self.output_prev_norm = None
+        del self.uuid_cache_diffs
+        self.uuid_cache_diffs = {}
+        self.total_steps_skipped = 0
+        self.state_metadata = None
+        return self
+
+    def clone(self):
+        return EasyCacheHolder(self.reuse_threshold, self.start_percent, self.end_percent, self.subsample_factor, self.offload_cache_diff, self.verbose)
+
+
+class EasyCacheNode(io.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="EasyCache",
+            display_name="EasyCache",
+            description="Native EasyCache implementation.",
+            category="advanced/debug/model",
+            is_experimental=True,
+            inputs=[
+                io.Model.Input("model", tooltip="The model to add EasyCache to."),
+                io.Float.Input("reuse_threshold", min=0.0, default=0.2, max=3.0, step=0.01, tooltip="The threshold for reusing cached steps."),
+                io.Float.Input("start_percent", min=0.0, default=0.15, max=1.0, step=0.01, tooltip="The relative sampling step to begin use of EasyCache."),
+                io.Float.Input("end_percent", min=0.0, default=0.95, max=1.0, step=0.01, tooltip="The relative sampling step to end use of EasyCache."),
+                io.Boolean.Input("verbose", default=False, tooltip="Whether to log verbose information."),
+            ],
+            outputs=[
+                io.Model.Output(tooltip="The model with EasyCache."),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, reuse_threshold: float, start_percent: float, end_percent: float, verbose: bool) -> io.NodeOutput:
+        model = model.clone()
+        model.model_options["transformer_options"]["easycache"] = EasyCacheHolder(reuse_threshold, start_percent, end_percent, subsample_factor=8, offload_cache_diff=False, verbose=verbose)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, "easycache", easycache_sample_wrapper)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.CALC_COND_BATCH, "easycache", easycache_calc_cond_batch_wrapper)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, "easycache", easycache_forward_wrapper)
+        return io.NodeOutput(model)
+
+
+class LazyCacheHolder:
+    def __init__(self, reuse_threshold: float, start_percent: float, end_percent: float, subsample_factor: int, offload_cache_diff: bool, verbose: bool=False):
+        self.name = "LazyCache"
+        self.reuse_threshold = reuse_threshold
+        self.start_percent = start_percent
+        self.end_percent = end_percent
+        self.subsample_factor = subsample_factor
+        self.offload_cache_diff = offload_cache_diff
+        self.verbose = verbose
+        # timestep values
+        self.start_t = 0.0
+        self.end_t = 0.0
+        # control values
+        self.relative_transformation_rate: float = None
+        self.cumulative_change_rate = 0.0
+        self.initial_step = True
+        # cache values
+        self.x_prev_subsampled: torch.Tensor = None
+        self.output_prev_subsampled: torch.Tensor = None
+        self.output_prev_norm: torch.Tensor = None
+        self.cache_diff: torch.Tensor = None
+        self.output_change_rates = []
+        self.approx_output_change_rates = []
+        self.total_steps_skipped = 0
+        self.state_metadata = None
+
+    def has_cache_diff(self) -> bool:
+        return self.cache_diff is not None
+
+    def is_past_end_timestep(self, timestep: float) -> bool:
+        return not (timestep[0] > self.end_t).item()
+
+    def should_do_easycache(self, timestep: float) -> bool:
+        return (timestep[0] <= self.start_t).item()
+
+    def has_x_prev_subsampled(self) -> bool:
+        return self.x_prev_subsampled is not None
+
+    def has_output_prev_subsampled(self) -> bool:
+        return self.output_prev_subsampled is not None
+
+    def has_output_prev_norm(self) -> bool:
+        return self.output_prev_norm is not None
+
+    def has_relative_transformation_rate(self) -> bool:
+        return self.relative_transformation_rate is not None
+
+    def prepare_timesteps(self, model_sampling):
+        self.start_t = model_sampling.percent_to_sigma(self.start_percent)
+        self.end_t = model_sampling.percent_to_sigma(self.end_percent)
+        return self
+
+    def subsample(self, x: torch.Tensor, clone: bool = True) -> torch.Tensor:
+        if self.subsample_factor > 1:
+            to_return = x[..., ::self.subsample_factor, ::self.subsample_factor]
+            if clone:
+                return to_return.clone()
+            return to_return
+        if clone:
+            return x.clone()
+        return x
+
+    def apply_cache_diff(self, x: torch.Tensor):
+        self.total_steps_skipped += 1
+        return x + self.cache_diff.to(x.device)
+
+    def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor):
+        self.cache_diff = output - x
+
+    def check_metadata(self, x: torch.Tensor) -> bool:
+        metadata = (x.device, x.dtype, x.shape)
+        if self.state_metadata is None:
+            self.state_metadata = metadata
+            return True
+        if metadata == self.state_metadata:
+            return True
+        logging.warn(f"{self.name} - Tensor shape, dtype or device changed, resetting state")
+        self.reset()
+        return False
+
+    def reset(self):
+        self.relative_transformation_rate = 0.0
+        self.cumulative_change_rate = 0.0
+        self.initial_step = True
+        self.output_change_rates = []
+        self.approx_output_change_rates = []
+        del self.cache_diff
+        self.cache_diff = None
+        del self.x_prev_subsampled
+        self.x_prev_subsampled = None
+        del self.output_prev_subsampled
+        self.output_prev_subsampled = None
+        del self.output_prev_norm
+        self.output_prev_norm = None
+        self.total_steps_skipped = 0
+        self.state_metadata = None
+        return self
+
+    def clone(self):
+        return LazyCacheHolder(self.reuse_threshold, self.start_percent, self.end_percent, self.subsample_factor, self.offload_cache_diff, self.verbose)
+
+class LazyCacheNode(io.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="LazyCache",
+            display_name="LazyCache",
+            description="A homebrew version of EasyCache - even 'easier' version of EasyCache to implement. Overall works worse than EasyCache, but better in some rare cases AND universal compatibility with everything in ComfyUI.",
+            category="advanced/debug/model",
+            is_experimental=True,
+            inputs=[
+                io.Model.Input("model", tooltip="The model to add LazyCache to."),
+                io.Float.Input("reuse_threshold", min=0.0, default=0.2, max=3.0, step=0.01, tooltip="The threshold for reusing cached steps."),
+                io.Float.Input("start_percent", min=0.0, default=0.15, max=1.0, step=0.01, tooltip="The relative sampling step to begin use of LazyCache."),
+                io.Float.Input("end_percent", min=0.0, default=0.95, max=1.0, step=0.01, tooltip="The relative sampling step to end use of LazyCache."),
+                io.Boolean.Input("verbose", default=False, tooltip="Whether to log verbose information."),
+            ],
+            outputs=[
+                io.Model.Output(tooltip="The model with LazyCache."),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, reuse_threshold: float, start_percent: float, end_percent: float, verbose: bool) -> io.NodeOutput:
+        model = model.clone()
+        model.model_options["transformer_options"]["easycache"] = LazyCacheHolder(reuse_threshold, start_percent, end_percent, subsample_factor=8, offload_cache_diff=False, verbose=verbose)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, "lazycache", easycache_sample_wrapper)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.PREDICT_NOISE, "lazycache", lazycache_predict_noise_wrapper)
+        return io.NodeOutput(model)
+
+
+class EasyCacheExtension(ComfyExtension):
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            EasyCacheNode,
+            LazyCacheNode,
+        ]
+
+def comfy_entrypoint():
+    return EasyCacheExtension()

comfy_extras/nodes_easysortblock.py

@@ -0,0 +1,503 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING, Union
+
+from comfy_api.latest import io, ComfyExtension
+import comfy.patcher_extension
+import logging
+import torch
+import math
+import comfy.model_patcher
+if TYPE_CHECKING:
+    from uuid import UUID
+
+
+
+def easysortblock_predict_noise_wrapper(executor, *args, **kwargs):
+    # get values from args
+    x: torch.Tensor = args[0]
+    timestep: float = args[1]
+    model_options: dict[str] = args[2]
+    easycache: EasySortblockHolder = model_options["transformer_options"]["easycache"]
+
+    # initialize predict_ratios
+    if easycache.initial_step:
+        sample_sigmas = model_options["transformer_options"]["sample_sigmas"]
+        relevant_sigmas = []
+        for i,sigma in enumerate(sample_sigmas):
+            if easycache.check_if_within_timesteps(sigma):
+                relevant_sigmas.append((i, sigma))
+        start_index = relevant_sigmas[0][0]
+        end_index = relevant_sigmas[-1][0]
+        easycache.predict_ratios = torch.linspace(easycache.start_predict_ratio, easycache.end_predict_ratio, end_index - start_index + 1)
+        easycache.predict_start_index = start_index
+
+    easycache.skip_current_step = False
+    if easycache.is_past_end_timestep(timestep):
+        return executor(*args, **kwargs)
+    # prepare next x_prev
+    next_x_prev = x
+    input_change = None
+    do_easycache = easycache.should_do_easycache(timestep)
+    if do_easycache:
+        easycache.check_metadata(x)
+        if easycache.has_x_prev_subsampled():
+            if easycache.has_x_prev_subsampled():
+                input_change = (easycache.subsample(x, clone=False) - easycache.x_prev_subsampled).flatten().abs().mean()
+            if easycache.has_output_prev_norm() and easycache.has_relative_transformation_rate():
+                approx_output_change_rate = (easycache.relative_transformation_rate * input_change) / easycache.output_prev_norm
+                easycache.cumulative_change_rate += approx_output_change_rate
+                if easycache.cumulative_change_rate < easycache.reuse_threshold:
+                    if easycache.verbose:
+                        logging.info(f"EasySortblock [verbose] - skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
+                    # other conds should also skip this step
+                    easycache.skip_current_step = True
+                    easycache.steps_skipped.append(easycache.step_count)
+                else:
+                    if easycache.verbose:
+                        logging.info(f"EasySortblock [verbose] - NOT skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
+                    easycache.cumulative_change_rate = 0.0
+    output: torch.Tensor = executor(*args, **kwargs)
+    if easycache.has_output_prev_norm():
+        output_change = (easycache.subsample(output, clone=False) - easycache.output_prev_subsampled).flatten().abs().mean()
+        if easycache.verbose:
+            output_change_rate = output_change / easycache.output_prev_norm
+            easycache.output_change_rates.append(output_change_rate.item())
+        if easycache.has_relative_transformation_rate():
+            approx_output_change_rate = (easycache.relative_transformation_rate * input_change) / easycache.output_prev_norm
+            easycache.approx_output_change_rates.append(approx_output_change_rate.item())
+            if easycache.verbose:
+                logging.info(f"EasySortblock [verbose] - approx_output_change_rate: {approx_output_change_rate}")
+        if input_change is not None:
+            easycache.relative_transformation_rate = output_change / input_change
+        if easycache.verbose:
+            logging.info(f"EasySortblock [verbose] - output_change_rate: {output_change_rate}")
+    easycache.x_prev_subsampled = easycache.subsample(next_x_prev)
+    easycache.output_prev_subsampled = easycache.subsample(output)
+    easycache.output_prev_norm = output.flatten().abs().mean()
+    if easycache.verbose:
+        logging.info(f"EasySortblock [verbose] - x_prev_subsampled: {easycache.x_prev_subsampled.shape}")
+
+    # increment step count
+    easycache.step_count += 1
+    easycache.initial_step = False
+
+    return output
+
+
+def easysortblock_outer_sample_wrapper(executor, *args, **kwargs):
+    """
+    This OUTER_SAMPLE wrapper makes sure EasySortblock is prepped for current run, and all memory usage is cleared at the end.
+    """
+    try:
+        guider = executor.class_obj
+        orig_model_options = guider.model_options
+        guider.model_options = comfy.model_patcher.create_model_options_clone(orig_model_options)
+        # clone and prepare timesteps
+        guider.model_options["transformer_options"]["easycache"] = guider.model_options["transformer_options"]["easycache"].clone().prepare_timesteps(guider.model_patcher.model.model_sampling)
+        easycache: EasySortblockHolder = guider.model_options['transformer_options']['easycache']
+        logging.info(f"{easycache.name} enabled - threshold: {easycache.reuse_threshold}, start_percent: {easycache.start_percent}, end_percent: {easycache.end_percent}")
+        return executor(*args, **kwargs)
+    finally:
+        easycache = guider.model_options['transformer_options']['easycache']
+        output_change_rates = easycache.output_change_rates
+        approx_output_change_rates = easycache.approx_output_change_rates
+        if easycache.verbose:
+            logging.info(f"{easycache.name} [verbose] - output_change_rates {len(output_change_rates)}: {output_change_rates}")
+            logging.info(f"{easycache.name} [verbose] - approx_output_change_rates {len(approx_output_change_rates)}: {approx_output_change_rates}")
+        total_steps = len(args[3])-1
+        logging.info(f"{easycache.name} - skipped {len(easycache.steps_skipped)}/{total_steps} steps")# ({total_steps/(total_steps-easycache.total_steps_skipped):.2f}x speedup).")
+        logging.info(f"{easycache.name} - skipped steps: {easycache.steps_skipped}")
+        easycache.reset()
+        guider.model_options = orig_model_options
+
+
+def model_forward_wrapper(executor, *args, **kwargs):
+    # TODO: make work with batches of conds
+    transformer_options: dict[str] = args[-1]
+    if not isinstance(transformer_options, dict):
+        transformer_options = kwargs.get("transformer_options")
+        if not transformer_options:
+            transformer_options = args[-2]
+    sigmas = transformer_options["sigmas"]
+    sb_holder: EasySortblockHolder = transformer_options["easycache"]
+
+    # if initial step, prepare everything for Sortblock
+    if sb_holder.initial_step:
+        logging.info(f"EasySortblock: inside model {executor.class_obj.__class__.__name__}")
+        # TODO: generalize for other models
+        # these won't stick around past this step; should store on sb_holder instead
+        logging.info(f"EasySortblock: preparing {len(executor.class_obj.double_blocks)} double blocks and {len(executor.class_obj.single_blocks)} single blocks")
+        if hasattr(executor.class_obj, "double_blocks"):
+            for block in executor.class_obj.double_blocks:
+                prepare_block(block, sb_holder)
+        if hasattr(executor.class_obj, "single_blocks"):
+            for block in executor.class_obj.single_blocks:
+                prepare_block(block, sb_holder)
+        if hasattr(executor.class_obj, "blocks"):
+            for block in executor.class_obj.block:
+                prepare_block(block, sb_holder)
+
+    if sb_holder.skip_current_step:
+        predict_index = max(0, sb_holder.step_count - sb_holder.predict_start_index)
+        predict_ratio = sb_holder.predict_ratios[predict_index]
+        logging.info(f"EasySortblock: skipping step {sb_holder.step_count}, predict_ratio: {predict_ratio}")
+        # reuse_ratio = 1.0 - predict_ratio
+        for block_type, blocks in sb_holder.blocks_per_type.items():
+            for block in blocks:
+                cache: BlockCache = block.__block_cache
+                cache.allowed_to_skip = False
+            sorted_blocks = sorted(blocks, key=lambda x: (x.__block_cache.consecutive_skipped_steps, x.__block_cache.prev_change_rate))
+            # for block in sorted_blocks:
+            #     pass
+            threshold_index = int(len(sorted_blocks) * predict_ratio)
+            # blocks with lower similarity are marked for recomputation
+            for block in sorted_blocks[:threshold_index]:
+                cache: BlockCache = block.__block_cache
+                cache.allowed_to_skip = True
+                logging.info(f"EasySortblock: skip block {block.__class__.__name__} - consecutive_skipped_steps: {block.__block_cache.consecutive_skipped_steps}, prev_change_rate: {block.__block_cache.prev_change_rate}, index: {block.__block_cache.block_index}")
+            not_skipped  = [block for block in blocks if not block.__block_cache.allowed_to_skip]
+            for block in not_skipped:
+                logging.info(f"EasySortblock: reco block {block.__class__.__name__} - consecutive_skipped_steps: {block.__block_cache.consecutive_skipped_steps}, prev_change_rate: {block.__block_cache.prev_change_rate}, index: {block.__block_cache.block_index}")
+            logging.info(f"EasySortblock: for {block_type}, selected {len(sorted_blocks[:threshold_index])} blocks for prediction and {len(sorted_blocks[threshold_index:])} blocks for recomputation")
+        # return executor(*args, **kwargs)
+    to_return = executor(*args, **kwargs)
+
+    return to_return
+
+
+
+
+
+def block_forward_factory(func, block):
+    def block_forward_wrapper(*args, **kwargs):
+        transformer_options: dict[str] = kwargs.get("transformer_options")
+        sigmas = transformer_options["sigmas"]
+        sb_holder: EasySortblockHolder = transformer_options["easycache"]
+        cache: BlockCache = block.__block_cache
+        # make sure stream count is properly set for this block
+        if sb_holder.initial_step:
+            sb_holder.add_to_blocks_per_type(block, transformer_options['block'][0])
+            cache.block_index = transformer_options['block'][1]
+            cache.stream_count = transformer_options['block'][2]
+
+        if sb_holder.is_past_end_timestep(sigmas):
+            return func(*args, **kwargs)
+        # do sortblock stuff
+        x = cache.get_next_x_prev(args, kwargs)
+        # prepare next_x_prev
+        next_x_prev = cache.get_next_x_prev(args, kwargs, clone=True)
+        input_change = None
+        do_sortblock = sb_holder.should_do_easycache(sigmas)
+        if do_sortblock:
+            # TODO: checkmetadata
+            if cache.has_x_prev_subsampled():
+                input_change = (cache.subsample(x, clone=False) - cache.x_prev_subsampled).flatten().abs().mean()
+            if cache.has_output_prev_norm() and cache.has_relative_transformation_rate():
+                approx_output_change_rate = (cache.relative_transformation_rate * input_change) / cache.output_prev_norm
+                cache.cumulative_change_rate += approx_output_change_rate
+                if cache.allowed_to_skip:
+                # if cache.cumulative_change_rate < sb_holder.reuse_threshold:
+                    # accumulate error + skip block
+                    # cache.want_to_skip = True
+                    # if cache.allowed_to_skip:
+                    cache.consecutive_skipped_steps += 1
+                    cache.prev_change_rate = approx_output_change_rate
+                    return cache.apply_cache_diff(x, sb_holder)
+                else:
+                    # reset error; NOT skipping block and recalculating
+                    cache.cumulative_change_rate = 0.0
+                    cache.prev_change_rate = approx_output_change_rate
+                    cache.want_to_skip = False
+                    cache.consecutive_skipped_steps = 0
+        # output_raw is expected to have cache.stream_count elements if count is greaater than 1 (double block, etc.)
+        output_raw: Union[torch.Tensor, tuple[torch.Tensor, ...]] = func(*args, **kwargs)
+        # if more than one stream from block, only use first one
+        if isinstance(output_raw, tuple):
+            output = output_raw[0]
+        else:
+            output = output_raw
+        if cache.has_output_prev_norm():
+            output_change = (cache.subsample(output, clone=False) - cache.output_prev_subsampled).flatten().abs().mean()
+            # if verbose in future
+            output_change_rate = output_change / cache.output_prev_norm
+            cache.output_change_rates.append(output_change_rate.item())
+            if cache.has_relative_transformation_rate():
+                approx_output_change_rate = (cache.relative_transformation_rate * input_change) / cache.output_prev_norm
+                cache.approx_output_change_rates.append(approx_output_change_rate.item())
+            if input_change is not None:
+                cache.relative_transformation_rate = output_change / input_change
+        # TODO: allow cache_diff to be offloaded
+        cache.update_cache_diff(output_raw, next_x_prev)
+        cache.x_prev_subsampled = cache.subsample(next_x_prev)
+        cache.output_prev_subsampled = cache.subsample(output)
+        cache.output_prev_norm = output.flatten().abs().mean()
+        return output_raw
+    return block_forward_wrapper
+
+def prepare_block(block, sb_holder: EasySortblockHolder, stream_count: int=1):
+    sb_holder.add_to_all_blocks(block)
+    block.__original_forward = block.forward
+    block.forward = block_forward_factory(block.__original_forward, block)
+    block.__block_cache = BlockCache(subsample_factor=sb_holder.subsample_factor, verbose=sb_holder.verbose)
+
+def clean_block(block):
+    block.forward = block.__original_forward
+    del block.__original_forward
+    del block.__block_cache
+
+class BlockCache:
+    def __init__(self, subsample_factor: int=8, verbose: bool=False):
+        self.subsample_factor = subsample_factor
+        self.verbose = verbose
+        self.stream_count = 1
+        self.block_index = 0
+        # control values
+        self.relative_transformation_rate: float = None
+        self.cumulative_change_rate = 0.0
+        self.prev_change_rate = 0.0
+        # cached values
+        self.x_prev_subsampled: torch.Tensor = None
+        self.output_prev_subsampled: torch.Tensor = None
+        self.output_prev_norm: torch.Tensor = None
+        self.cache_diff: list[torch.Tensor] = []
+        self.output_change_rates = []
+        self.approx_output_change_rates = []
+        self.steps_skipped: list[int] = []
+        self.consecutive_skipped_steps = 0
+        # self.state_metadata = None
+        self.want_to_skip = False
+        self.allowed_to_skip = False
+
+    def has_cache_diff(self) -> bool:
+        return self.cache_diff[0] is not None
+
+    def has_x_prev_subsampled(self) -> bool:
+        return self.x_prev_subsampled is not None
+
+    def has_output_prev_subsampled(self) -> bool:
+        return self.output_prev_subsampled is not None
+
+    def has_output_prev_norm(self) -> bool:
+        return self.output_prev_norm is not None
+
+    def has_relative_transformation_rate(self) -> bool:
+        return self.relative_transformation_rate is not None
+
+    def get_next_x_prev(self, d_args: tuple[torch.Tensor, ...], d_kwargs: dict[str, torch.Tensor], clone: bool=False) -> tuple[torch.Tensor, ...]:
+        if self.stream_count == 1:
+            if clone:
+                return d_args[0].clone()
+            return d_args[0]
+        keys = list(d_kwargs.keys())[:self.stream_count]
+        orig_inputs = []
+        for key in keys:
+            if clone:
+                orig_inputs.append(d_kwargs[key].clone())
+            else:
+                orig_inputs.append(d_kwargs[key])
+        return tuple(orig_inputs)
+
+    def subsample(self, x: Union[torch.Tensor, tuple[torch.Tensor, ...]], clone: bool = True) -> torch.Tensor:
+        # subsample only the first compoenent
+        if isinstance(x, tuple):
+            return self.subsample(x[0], clone)
+        if self.subsample_factor > 1:
+            to_return = x[..., ::self.subsample_factor, ::self.subsample_factor]
+            if clone:
+                return to_return.clone()
+            return to_return
+        if clone:
+            return x.clone()
+        return x
+
+    def apply_cache_diff(self, x: Union[torch.Tensor, tuple[torch.Tensor, ...]], sb_holder: EasySortblockHolder):
+        self.steps_skipped.append(sb_holder.step_count)
+        if not isinstance(x, tuple):
+            x = (x, )
+        to_return = tuple([x[i] + self.cache_diff[i] for i in range(self.stream_count)])
+        if len(to_return) == 1:
+            return to_return[0]
+        return to_return
+
+    def update_cache_diff(self, output_raw: Union[torch.Tensor, tuple[torch.Tensor, ...]], x: Union[torch.Tensor, tuple[torch.Tensor, ...]]):
+        if not isinstance(output_raw, tuple):
+            output_raw = (output_raw, )
+        if not isinstance(x, tuple):
+            x = (x, )
+        self.cache_diff = tuple([output_raw[i] - x[i] for i in range(self.stream_count)])
+
+    def reset(self):
+        self.relative_transformation_rate = 0.0
+        self.cumulative_change_rate = 0.0
+        self.prev_change_rate = 0.0
+        self.x_prev_subsampled = None
+        self.output_prev_subsampled = None
+        self.output_prev_norm = None
+        self.cache_diff = []
+        self.output_change_rates = []
+        self.approx_output_change_rates = []
+        self.steps_skipped = []
+        self.consecutive_skipped_steps = 0
+        self.want_to_skip = False
+        self.allowed_to_skip = False
+        return self
+
+
+class EasySortblockHolder:
+    def __init__(self, reuse_threshold: float, start_predict_ratio: float, end_predict_ratio: float, max_skipped_steps: int,
+                       start_percent: float, end_percent: float, subsample_factor: int, verbose: bool=False):
+        self.name = "EasySortblock"
+        self.reuse_threshold = reuse_threshold
+        self.start_predict_ratio = start_predict_ratio
+        self.end_predict_ratio = end_predict_ratio
+        self.max_skipped_steps = max_skipped_steps
+        self.start_percent = start_percent
+        self.end_percent = end_percent
+        self.subsample_factor = subsample_factor
+        self.verbose = verbose
+        # timestep values
+        self.start_t = 0.0
+        self.end_t = 0.0
+        # control values
+        self.relative_transformation_rate: float = None
+        self.cumulative_change_rate = 0.0
+        self.initial_step = True
+        self.step_count = 0
+        self.predict_ratios = []
+        self.skip_current_step = False
+        self.predict_start_index = 0
+        # cache values
+        self.x_prev_subsampled: torch.Tensor = None
+        self.output_prev_subsampled: torch.Tensor = None
+        self.output_prev_norm: torch.Tensor = None
+        self.steps_skipped: list[int] = []
+        self.output_change_rates = []
+        self.approx_output_change_rates = []
+        self.state_metadata = None
+        self.all_blocks = []
+        self.blocks_per_type = {}
+
+    def add_to_all_blocks(self, block):
+        self.all_blocks.append(block)
+
+    def add_to_blocks_per_type(self, block, block_type: str):
+        self.blocks_per_type.setdefault(block_type, []).append(block)
+
+    def is_past_end_timestep(self, timestep: float) -> bool:
+        return not (timestep[0] > self.end_t).item()
+
+    def should_do_easycache(self, timestep: float) -> bool:
+        return (timestep[0] <= self.start_t).item()
+
+    def check_if_within_timesteps(self, timestep: Union[float, torch.Tensor]) -> bool:
+        return (timestep <= self.start_t).item() and (timestep > self.end_t).item()
+
+    def has_x_prev_subsampled(self) -> bool:
+        return self.x_prev_subsampled is not None
+
+    def has_output_prev_subsampled(self) -> bool:
+        return self.output_prev_subsampled is not None
+
+    def has_output_prev_norm(self) -> bool:
+        return self.output_prev_norm is not None
+
+    def has_relative_transformation_rate(self) -> bool:
+        return self.relative_transformation_rate is not None
+
+    def prepare_timesteps(self, model_sampling):
+        self.start_t = model_sampling.percent_to_sigma(self.start_percent)
+        self.end_t = model_sampling.percent_to_sigma(self.end_percent)
+        return self
+
+    def subsample(self, x: torch.Tensor, clone: bool = True) -> torch.Tensor:
+        if self.subsample_factor > 1:
+            to_return = x[..., ::self.subsample_factor, ::self.subsample_factor]
+            if clone:
+                return to_return.clone()
+            return to_return
+        if clone:
+            return x.clone()
+        return x
+
+    def check_metadata(self, x: torch.Tensor) -> bool:
+        metadata = (x.device, x.dtype, x.shape)
+        if self.state_metadata is None:
+            self.state_metadata = metadata
+            return True
+        if metadata == self.state_metadata:
+            return True
+        logging.warning(f"{self.name} - Tensor shape, dtype or device changed, resetting state")
+        self.reset()
+        return False
+
+    def reset(self):
+        logging.info(f"EasySortblock: resetting {len(self.all_blocks)} blocks")
+        for block in self.all_blocks:
+            clean_block(block)
+        self.relative_transformation_rate = 0.0
+        self.cumulative_change_rate = 0.0
+        self.initial_step = True
+        self.step_count = 0
+        self.predict_ratios = []
+        self.skip_current_step = False
+        self.predict_start_index = 0
+        self.x_prev_subsampled = None
+        self.output_prev_subsampled = None
+        self.output_prev_norm = None
+        self.steps_skipped = []
+        self.output_change_rates = []
+        self.approx_output_change_rates = []
+        self.state_metadata = None
+        self.all_blocks = []
+        self.blocks_per_type = {}
+        return self
+
+    def clone(self):
+        return EasySortblockHolder(self.reuse_threshold, self.start_predict_ratio, self.end_predict_ratio, self.max_skipped_steps,
+                                   self.start_percent, self.end_percent, self.subsample_factor, self.verbose)
+
+class EasySortblockScaledNode(io.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="EasySortblockScaled",
+            display_name="EasySortblockScaled",
+            description="A homebrew version of EasyCache - even 'easier' version of EasyCache to implement. Overall works worse than EasyCache, but better in some rare cases AND universal compatibility with everything in ComfyUI.",
+            category="advanced/debug/model",
+            is_experimental=True,
+            inputs=[
+                io.Model.Input("model", tooltip="The model to add Sortblock to."),
+                io.Float.Input("reuse_threshold", min=0.0, default=0.2, max=3.0, step=0.01, tooltip="The threshold for reusing cached steps."),
+                io.Float.Input("start_predict_ratio", min=0.0, default=0.2, max=1.0, step=0.01, tooltip="The ratio of blocks to predict."),
+                io.Float.Input("end_predict_ratio", min=0.0, default=0.9, max=1.0, step=0.01, tooltip="The ratio of blocks to predict."),
+                io.Int.Input("policy_refresh_interval", min=3, default=5, max=100, step=1, tooltip="The interval at which to refresh the policy."),
+                io.Float.Input("start_percent", min=0.0, default=0.15, max=1.0, step=0.01, tooltip="The relative sampling step to begin use of Sortblock."),
+                io.Float.Input("end_percent", min=0.0, default=0.95, max=1.0, step=0.01, tooltip="The relative sampling step to end use of Sortblock."),
+                io.Boolean.Input("verbose", default=False, tooltip="Whether to log verbose information."),
+            ],
+            outputs=[
+                io.Model.Output(tooltip="The model with Sortblock."),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, reuse_threshold: float, start_predict_ratio: float, end_predict_ratio: float, policy_refresh_interval: int, start_percent: float, end_percent: float, verbose: bool) -> io.NodeOutput:
+        # TODO: check for specific flavors of supported models
+        model = model.clone()
+        model.model_options["transformer_options"]["easycache"] = EasySortblockHolder(reuse_threshold, start_predict_ratio, end_predict_ratio, policy_refresh_interval, start_percent, end_percent, subsample_factor=8, verbose=verbose)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.PREDICT_NOISE, "sortblock", easysortblock_predict_noise_wrapper)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, "sortblock", easysortblock_outer_sample_wrapper)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, "sortblock", model_forward_wrapper)
+        return io.NodeOutput(model)
+
+
+class EasySortblockExtension(ComfyExtension):
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            # EasySortblockNode,
+            EasySortblockScaledNode,
+        ]
+
+def comfy_entrypoint():
+    return EasySortblockExtension()
+

comfy_extras/nodes_flux.py

@@ -100,9 +100,28 @@ class FluxKontextImageScale:
         return (image, )
 
 
+class FluxKontextMultiReferenceLatentMethod:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "conditioning": ("CONDITIONING", ),
+            "reference_latents_method": (("offset", "index"), ),
+            }}
+
+    RETURN_TYPES = ("CONDITIONING",)
+    FUNCTION = "append"
+    EXPERIMENTAL = True
+
+    CATEGORY = "advanced/conditioning/flux"
+
+    def append(self, conditioning, reference_latents_method):
+        c = node_helpers.conditioning_set_values(conditioning, {"reference_latents_method": reference_latents_method})
+        return (c, )
+
 NODE_CLASS_MAPPINGS = {
     "CLIPTextEncodeFlux": CLIPTextEncodeFlux,
     "FluxGuidance": FluxGuidance,
     "FluxDisableGuidance": FluxDisableGuidance,
     "FluxKontextImageScale": FluxKontextImageScale,
+    "FluxKontextMultiReferenceLatentMethod": FluxKontextMultiReferenceLatentMethod,
 }

comfy_extras/nodes_latent.py

@@ -1,6 +1,7 @@
 import comfy.utils
 import comfy_extras.nodes_post_processing
 import torch
+import nodes
 
 
 def reshape_latent_to(target_shape, latent, repeat_batch=True):
@@ -105,6 +106,73 @@ class LatentInterpolate:
         samples_out["samples"] = st * (m1 * ratio + m2 * (1.0 - ratio))
         return (samples_out,)
 
+class LatentConcat:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "samples1": ("LATENT",), "samples2": ("LATENT",), "dim": (["x", "-x", "y", "-y", "t", "-t"], )}}
+
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "op"
+
+    CATEGORY = "latent/advanced"
+
+    def op(self, samples1, samples2, dim):
+        samples_out = samples1.copy()
+
+        s1 = samples1["samples"]
+        s2 = samples2["samples"]
+        s2 = comfy.utils.repeat_to_batch_size(s2, s1.shape[0])
+
+        if "-" in dim:
+            c = (s2, s1)
+        else:
+            c = (s1, s2)
+
+        if "x" in dim:
+            dim = -1
+        elif "y" in dim:
+            dim = -2
+        elif "t" in dim:
+            dim = -3
+
+        samples_out["samples"] = torch.cat(c, dim=dim)
+        return (samples_out,)
+
+class LatentCut:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"samples": ("LATENT",),
+                             "dim": (["x", "y", "t"], ),
+                             "index": ("INT", {"default": 0, "min": -nodes.MAX_RESOLUTION, "max": nodes.MAX_RESOLUTION, "step": 1}),
+                             "amount": ("INT", {"default": 1, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 1})}}
+
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "op"
+
+    CATEGORY = "latent/advanced"
+
+    def op(self, samples, dim, index, amount):
+        samples_out = samples.copy()
+
+        s1 = samples["samples"]
+
+        if "x" in dim:
+            dim = s1.ndim - 1
+        elif "y" in dim:
+            dim = s1.ndim - 2
+        elif "t" in dim:
+            dim = s1.ndim - 3
+
+        if index >= 0:
+            index = min(index, s1.shape[dim] - 1)
+            amount = min(s1.shape[dim] - index, amount)
+        else:
+            index = max(index, -s1.shape[dim])
+            amount = min(-index, amount)
+
+        samples_out["samples"] = torch.narrow(s1, dim, index, amount)
+        return (samples_out,)
+
 class LatentBatch:
     @classmethod
     def INPUT_TYPES(s):
@@ -279,6 +347,8 @@ NODE_CLASS_MAPPINGS = {
     "LatentSubtract": LatentSubtract,
     "LatentMultiply": LatentMultiply,
     "LatentInterpolate": LatentInterpolate,
+    "LatentConcat": LatentConcat,
+    "LatentCut": LatentCut,
     "LatentBatch": LatentBatch,
     "LatentBatchSeedBehavior": LatentBatchSeedBehavior,
     "LatentApplyOperation": LatentApplyOperation,

comfy_extras/nodes_lt.py

@@ -166,7 +166,7 @@ class LTXVAddGuide:
         negative = self.add_keyframe_index(negative, frame_idx, guiding_latent, scale_factors)
 
         mask = torch.full(
-            (noise_mask.shape[0], 1, guiding_latent.shape[2], 1, 1),
+            (noise_mask.shape[0], 1, guiding_latent.shape[2], noise_mask.shape[3], noise_mask.shape[4]),
             1.0 - strength,
             dtype=noise_mask.dtype,
             device=noise_mask.device,

comfy_extras/nodes_model_patch.py

@@ -0,0 +1,163 @@
+import torch
+import folder_paths
+import comfy.utils
+import comfy.ops
+import comfy.model_management
+import comfy.ldm.common_dit
+import comfy.latent_formats
+
+
+class BlockWiseControlBlock(torch.nn.Module):
+    # [linear, gelu, linear]
+    def __init__(self, dim: int = 3072, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.x_rms = operations.RMSNorm(dim, eps=1e-6)
+        self.y_rms = operations.RMSNorm(dim, eps=1e-6)
+        self.input_proj = operations.Linear(dim, dim)
+        self.act = torch.nn.GELU()
+        self.output_proj = operations.Linear(dim, dim)
+
+    def forward(self, x, y):
+        x, y = self.x_rms(x), self.y_rms(y)
+        x = self.input_proj(x + y)
+        x = self.act(x)
+        x = self.output_proj(x)
+        return x
+
+
+class QwenImageBlockWiseControlNet(torch.nn.Module):
+    def __init__(
+        self,
+        num_layers: int = 60,
+        in_dim: int = 64,
+        additional_in_dim: int = 0,
+        dim: int = 3072,
+        device=None, dtype=None, operations=None
+    ):
+        super().__init__()
+        self.additional_in_dim = additional_in_dim
+        self.img_in = operations.Linear(in_dim + additional_in_dim, dim, device=device, dtype=dtype)
+        self.controlnet_blocks = torch.nn.ModuleList(
+            [
+                BlockWiseControlBlock(dim, device=device, dtype=dtype, operations=operations)
+                for _ in range(num_layers)
+            ]
+        )
+
+    def process_input_latent_image(self, latent_image):
+        latent_image[:, :16] = comfy.latent_formats.Wan21().process_in(latent_image[:, :16])
+        patch_size = 2
+        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(latent_image, (1, patch_size, patch_size))
+        orig_shape = hidden_states.shape
+        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
+        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5)
+        hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
+        return self.img_in(hidden_states)
+
+    def control_block(self, img, controlnet_conditioning, block_id):
+        return self.controlnet_blocks[block_id](img, controlnet_conditioning)
+
+
+class ModelPatchLoader:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "name": (folder_paths.get_filename_list("model_patches"), ),
+                              }}
+    RETURN_TYPES = ("MODEL_PATCH",)
+    FUNCTION = "load_model_patch"
+    EXPERIMENTAL = True
+
+    CATEGORY = "advanced/loaders"
+
+    def load_model_patch(self, name):
+        model_patch_path = folder_paths.get_full_path_or_raise("model_patches", name)
+        sd = comfy.utils.load_torch_file(model_patch_path, safe_load=True)
+        dtype = comfy.utils.weight_dtype(sd)
+        # TODO: this node will work with more types of model patches
+        additional_in_dim = sd["img_in.weight"].shape[1] - 64
+        model = QwenImageBlockWiseControlNet(additional_in_dim=additional_in_dim, device=comfy.model_management.unet_offload_device(), dtype=dtype, operations=comfy.ops.manual_cast)
+        model.load_state_dict(sd)
+        model = comfy.model_patcher.ModelPatcher(model, load_device=comfy.model_management.get_torch_device(), offload_device=comfy.model_management.unet_offload_device())
+        return (model,)
+
+
+class DiffSynthCnetPatch:
+    def __init__(self, model_patch, vae, image, strength, mask=None):
+        self.model_patch = model_patch
+        self.vae = vae
+        self.image = image
+        self.strength = strength
+        self.mask = mask
+        self.encoded_image = model_patch.model.process_input_latent_image(self.encode_latent_cond(image))
+        self.encoded_image_size = (image.shape[1], image.shape[2])
+
+    def encode_latent_cond(self, image):
+        latent_image = self.vae.encode(image)
+        if self.model_patch.model.additional_in_dim > 0:
+            if self.mask is None:
+                mask_ = torch.ones_like(latent_image)[:, :self.model_patch.model.additional_in_dim // 4]
+            else:
+                mask_ = comfy.utils.common_upscale(self.mask.mean(dim=1, keepdim=True), latent_image.shape[-1], latent_image.shape[-2], "bilinear", "none")
+
+            return torch.cat([latent_image, mask_], dim=1)
+        else:
+            return latent_image
+
+    def __call__(self, kwargs):
+        x = kwargs.get("x")
+        img = kwargs.get("img")
+        block_index = kwargs.get("block_index")
+        spacial_compression = self.vae.spacial_compression_encode()
+        if self.encoded_image is None or self.encoded_image_size != (x.shape[-2] * spacial_compression, x.shape[-1] * spacial_compression):
+            image_scaled = comfy.utils.common_upscale(self.image.movedim(-1, 1), x.shape[-1] * spacial_compression, x.shape[-2] * spacial_compression, "area", "center")
+            loaded_models = comfy.model_management.loaded_models(only_currently_used=True)
+            self.encoded_image = self.model_patch.model.process_input_latent_image(self.encode_latent_cond(image_scaled.movedim(1, -1)))
+            self.encoded_image_size = (image_scaled.shape[-2], image_scaled.shape[-1])
+            comfy.model_management.load_models_gpu(loaded_models)
+
+        img[:, :self.encoded_image.shape[1]] += (self.model_patch.model.control_block(img[:, :self.encoded_image.shape[1]], self.encoded_image.to(img.dtype), block_index) * self.strength)
+        kwargs['img'] = img
+        return kwargs
+
+    def to(self, device_or_dtype):
+        if isinstance(device_or_dtype, torch.device):
+            self.encoded_image = self.encoded_image.to(device_or_dtype)
+        return self
+
+    def models(self):
+        return [self.model_patch]
+
+class QwenImageDiffsynthControlnet:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "model_patch": ("MODEL_PATCH",),
+                              "vae": ("VAE",),
+                              "image": ("IMAGE",),
+                              "strength": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}),
+                              },
+                "optional": {"mask": ("MASK",)}}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "diffsynth_controlnet"
+    EXPERIMENTAL = True
+
+    CATEGORY = "advanced/loaders/qwen"
+
+    def diffsynth_controlnet(self, model, model_patch, vae, image, strength, mask=None):
+        model_patched = model.clone()
+        image = image[:, :, :, :3]
+        if mask is not None:
+            if mask.ndim == 3:
+                mask = mask.unsqueeze(1)
+            if mask.ndim == 4:
+                mask = mask.unsqueeze(2)
+            mask = 1.0 - mask
+
+        model_patched.set_model_double_block_patch(DiffSynthCnetPatch(model_patch, vae, image, strength, mask))
+        return (model_patched,)
+
+
+NODE_CLASS_MAPPINGS = {
+    "ModelPatchLoader": ModelPatchLoader,
+    "QwenImageDiffsynthControlnet": QwenImageDiffsynthControlnet,
+}

comfy_extras/nodes_qwen.py

@@ -0,0 +1,48 @@
+import node_helpers
+import comfy.utils
+import math
+
+
+class TextEncodeQwenImageEdit:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "clip": ("CLIP", ),
+            "prompt": ("STRING", {"multiline": True, "dynamicPrompts": True}),
+            },
+            "optional": {"vae": ("VAE", ),
+                         "image": ("IMAGE", ),}}
+
+    RETURN_TYPES = ("CONDITIONING",)
+    FUNCTION = "encode"
+
+    CATEGORY = "advanced/conditioning"
+
+    def encode(self, clip, prompt, vae=None, image=None):
+        ref_latent = None
+        if image is None:
+            images = []
+        else:
+            samples = image.movedim(-1, 1)
+            total = int(1024 * 1024)
+
+            scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2]))
+            width = round(samples.shape[3] * scale_by)
+            height = round(samples.shape[2] * scale_by)
+
+            s = comfy.utils.common_upscale(samples, width, height, "area", "disabled")
+            image = s.movedim(1, -1)
+            images = [image[:, :, :, :3]]
+            if vae is not None:
+                ref_latent = vae.encode(image[:, :, :, :3])
+
+        tokens = clip.tokenize(prompt, images=images)
+        conditioning = clip.encode_from_tokens_scheduled(tokens)
+        if ref_latent is not None:
+            conditioning = node_helpers.conditioning_set_values(conditioning, {"reference_latents": [ref_latent]}, append=True)
+        return (conditioning, )
+
+
+NODE_CLASS_MAPPINGS = {
+    "TextEncodeQwenImageEdit": TextEncodeQwenImageEdit,
+}

comfy_extras/nodes_sortblock.py

@@ -0,0 +1,462 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING, Union
+
+from comfy_api.latest import io, ComfyExtension
+import comfy.patcher_extension
+import logging
+import torch
+import math
+import comfy.model_patcher
+if TYPE_CHECKING:
+    from uuid import UUID
+
+
+def prepare_noise_wrapper(executor, *args, **kwargs):
+    try:
+        transformer_options: dict[str] = args[2]["transformer_options"]
+        sb_holder: SortblockHolder = transformer_options["sortblock"]
+        if sb_holder.initial_step:
+            sample_sigmas = transformer_options["sample_sigmas"]
+            relevant_sigmas = []
+            # find start and end steps, then use to interpolate between start and end predict ratios
+            for i,sigma in enumerate(sample_sigmas):
+                if sb_holder.check_if_within_timesteps(sigma):
+                    relevant_sigmas.append((i, sigma))
+            start_index = relevant_sigmas[0][0]
+            end_index = relevant_sigmas[-1][0]
+            sb_holder.predict_ratios = torch.linspace(sb_holder.start_predict_ratio, sb_holder.end_predict_ratio, end_index - start_index + 1)
+            sb_holder.predict_start_index = start_index
+
+        return executor(*args, **kwargs)
+    finally:
+        transformer_options: dict[str] = args[2]["transformer_options"]
+        sb_holder: SortblockHolder = transformer_options["sortblock"]
+        sb_holder.step_count += 1
+        if sb_holder.should_do_sortblock():
+            sb_holder.active_steps += 1
+
+
+def outer_sample_wrapper(executor, *args, **kwargs):
+    try:
+        logging.info("Sortblock: inside outer_sample!")
+        guider = executor.class_obj
+        orig_model_options = guider.model_options
+        guider.model_options = comfy.model_patcher.create_model_options_clone(orig_model_options)
+        # clone and prepare timesteps
+        sb_holder = guider.model_options["transformer_options"]["sortblock"]
+        guider.model_options["transformer_options"]["sortblock"] = sb_holder.clone().prepare_timesteps(guider.model_patcher.model.model_sampling)
+        sb_holder: SortblockHolder = guider.model_options["transformer_options"]["sortblock"]
+        logging.info(f"Sortblock: enabled - threshold: {sb_holder.start_predict_ratio}, start_percent: {sb_holder.start_percent}, end_percent: {sb_holder.end_percent}")
+        return executor(*args, **kwargs)
+    finally:
+        sb_holder = guider.model_options["transformer_options"]["sortblock"]
+        logging.info(f"Sortblock: final step count: {sb_holder.step_count}")
+        sb_holder.reset()
+        guider.model_options = orig_model_options
+
+
+def model_forward_wrapper(executor, *args, **kwargs):
+    # TODO: make work with batches of conds
+    transformer_options: dict[str] = args[-1]
+    if not isinstance(transformer_options, dict):
+        transformer_options = kwargs.get("transformer_options")
+        if not transformer_options:
+            transformer_options = args[-2]
+    sigmas = transformer_options["sigmas"]
+    sb_holder: SortblockHolder = transformer_options["sortblock"]
+    sb_holder.update_should_do_sortblock(sigmas)
+
+    # if initial step, prepare everything for Sortblock
+    if sb_holder.initial_step:
+        logging.info(f"Sortblock: inside model {executor.class_obj.__class__.__name__}")
+        # TODO: generalize for other models
+        # these won't stick around past this step; should store on sb_holder instead
+        logging.info(f"Sortblock: preparing {len(executor.class_obj.double_blocks)} double blocks and {len(executor.class_obj.single_blocks)} single blocks")
+        if hasattr(executor.class_obj, "double_blocks"):
+            for block in executor.class_obj.double_blocks:
+                prepare_block(block, sb_holder)
+        if hasattr(executor.class_obj, "single_blocks"):
+            for block in executor.class_obj.single_blocks:
+                prepare_block(block, sb_holder)
+        if hasattr(executor.class_obj, "blocks"):
+            for block in executor.class_obj.block:
+                prepare_block(block, sb_holder)
+
+    # when 0: Initialization(1)
+    if sb_holder.step_modulus == 0:
+        logging.info(f"Sortblock: for step {sb_holder.step_count}, all blocks are marked for recomputation")
+        # all features are computed, input-outputs changes for all DiT blocks are stored for relative step 'k'
+        sb_holder.activated_steps.append(sb_holder.step_count)
+        for block in sb_holder.all_blocks:
+            cache: BlockCache = block.__block_cache
+            cache.mark_recompute()
+
+    # all block operations are performed in forward pass of model
+    to_return = executor(*args, **kwargs)
+
+    # when 1: Select DiT blocks(4)
+    if sb_holder.step_modulus == 1:
+        predict_index = max(0, sb_holder.step_count - sb_holder.predict_start_index)
+        predict_ratio = sb_holder.predict_ratios[predict_index]
+        logging.info(f"Sortblock: for step {sb_holder.step_count}, selecting blocks for recomputation and prediction, predict_ratio: {predict_ratio}")
+        reuse_ratio = 1.0 - predict_ratio
+        for block_type, blocks in sb_holder.blocks_per_type.items():
+            sorted_blocks = sorted(blocks, key=lambda x: x.__block_cache.cosine_similarity)
+            threshold_index = int(len(sorted_blocks) * reuse_ratio)
+            # blocks with lower similarity are marked for recomputation
+            for block in sorted_blocks[:threshold_index]:
+                cache: BlockCache = block.__block_cache
+                cache.mark_recompute()
+            # blocks with higher similarity are marked for prediction
+            for block in sorted_blocks[threshold_index:]:
+                cache: BlockCache = block.__block_cache
+                cache.mark_predict()
+            logging.info(f"Sortblock: for {block_type}, selected {len(sorted_blocks[:threshold_index])} blocks for recomputation and {len(sorted_blocks[threshold_index:])} blocks for prediction")
+
+    if sb_holder.initial_step:
+        sb_holder.initial_step = False
+    return to_return
+
+def block_forward_factory(func, block):
+    def block_forward_wrapper(*args, **kwargs):
+        transformer_options: dict[str] = kwargs.get("transformer_options")
+        sb_holder: SortblockHolder = transformer_options["sortblock"]
+        cache: BlockCache = block.__block_cache
+        # make sure stream count is properly set for this block
+        if sb_holder.initial_step:
+            sb_holder.add_to_blocks_per_type(block, transformer_options['block'][0])
+            cache.block_index = transformer_options['block'][1]
+            cache.stream_count = transformer_options['block'][2]
+        # do sortblock stuff
+        if cache.recompute and sb_holder.step_modulus != 1:
+            # clone relevant inputs
+            orig_inputs = cache.get_orig_inputs(args, kwargs, clone=True)
+            # get block outputs
+            # NOTE: output_raw is expected to have cache.stream_count elements if count is greaater than 1 (double block, etc.)
+            if cache.stream_count == 1:
+                zzz = 10
+            output_raw: Union[torch.Tensor, tuple[torch.Tensor, ...]] = func(*args, **kwargs)
+            # perform derivative approximation;
+            cache.derivative_approximation(sb_holder, output_raw, orig_inputs)
+            # if step_modulus is 0, input-output changes for DiT block are stored
+            if sb_holder.step_modulus == 0:
+                cache.cache_previous_residual(output_raw, orig_inputs)
+        else:
+            # if not to recompute, predict features for current timestep
+            orig_inputs = cache.get_orig_inputs(args, kwargs, clone=False)
+            # when 1: Linear Prediction(2)
+            # if step_modulus is 1, store block residuals as 'current' after applying taylor_formula
+            if sb_holder.step_modulus == 1:
+                cache.cache_current_residual(sb_holder)
+            # based on features computed in last timestep, all features for current timestep are predicted using Eq. 4,
+            # input-output changes for all DiT blocks are stored for relative step 'k+1'
+            output_raw = cache.apply_linear_prediction(sb_holder, orig_inputs)
+
+        # when 1: Identify Changes(3)
+        if sb_holder.step_modulus == 1:
+            # based on features computed in last timestep, all features for current timestep are predicted using Eq. 4,
+            # input-output changes for all DiT blocks are stored for relative step 'k+1'
+            cache.calculate_cosine_similarity()
+
+        # return output_raw
+        return output_raw
+    return block_forward_wrapper
+
+
+def perform_sortblock(blocks: list):
+    ...
+
+def prepare_block(block, sb_holder: SortblockHolder, stream_count: int=1):
+    sb_holder.add_to_all_blocks(block)
+    block.__original_forward = block.forward
+    block.forward = block_forward_factory(block.__original_forward, block)
+    block.__block_cache = BlockCache(subsample_factor=sb_holder.subsample_factor, verbose=sb_holder.verbose)
+
+def clean_block(block):
+    block.forward = block.__original_forward
+    del block.__original_forward
+    del block.__block_cache
+
+def subsample(x: torch.Tensor, factor: int, clone: bool=True) -> torch.Tensor:
+    if factor > 1:
+        to_return = x[..., ::factor, ::factor]
+        if clone:
+            return to_return.clone()
+        return to_return
+    if clone:
+        return x.clone()
+    return x
+
+class BlockCache:
+    def __init__(self, subsample_factor: int=8, verbose: bool=False):
+        self.subsample_factor = subsample_factor
+        self.verbose = verbose
+        self.stream_count = 1
+        self.recompute = False
+        self.block_index = 0
+        # cached values
+        self.previous_residual_subsampled: torch.Tensor = None
+        self.current_residual_subsampled: torch.Tensor = None
+        self.cosine_similarity: float = None
+        self.previous_taylor_factors: dict[int, torch.Tensor] = {}
+        self.current_taylor_factors: dict[int, torch.Tensor] = {}
+
+    def mark_recompute(self):
+        self.recompute = True
+
+    def mark_predict(self):
+        self.recompute = False
+
+    def cache_previous_residual(self, output_raw: Union[torch.Tensor, tuple[torch.Tensor, ...]], orig_inputs: Union[torch.Tensor, tuple[torch.Tensor, ...]]):
+        if isinstance(output_raw, tuple):
+            output_raw = output_raw[0]
+        if isinstance(orig_inputs, tuple):
+            orig_inputs = orig_inputs[0]
+        del self.previous_residual_subsampled
+        self.previous_residual_subsampled = subsample(output_raw - orig_inputs, self.subsample_factor, clone=True)
+
+    def cache_current_residual(self, sb_holder: SortblockHolder):
+        del self.current_residual_subsampled
+        self.current_residual_subsampled = subsample(self.use_taylor_formula(sb_holder)[0], self.subsample_factor, clone=True)
+
+    def get_orig_inputs(self, d_args: tuple, d_kwargs: dict, clone: bool=True) -> tuple[torch.Tensor, ...]:
+        if self.stream_count == 1:
+            if clone:
+                return d_args[0].clone()
+            return d_args[0]
+        keys = list(d_kwargs.keys())[:self.stream_count]
+        orig_inputs = []
+        for key in keys:
+            if clone:
+                orig_inputs.append(d_kwargs[key].clone())
+            else:
+                orig_inputs.append(d_kwargs[key])
+        return tuple(orig_inputs)
+
+    def apply_linear_prediction(self, sb_holder: SortblockHolder, orig_inputs: Union[torch.Tensor, tuple[torch.Tensor, ...]]) -> None:
+        drop_tuple = False
+        if not isinstance(orig_inputs, tuple):
+            orig_inputs = (orig_inputs,)
+            drop_tuple = True
+        taylor_results = self.use_taylor_formula(sb_holder)
+        for output, taylor_result in zip(orig_inputs, taylor_results):
+            if output.shape != taylor_result.shape:
+                zzz = 10
+            output += taylor_result
+        if drop_tuple:
+            orig_inputs = orig_inputs[0]
+        return orig_inputs
+
+    def calculate_cosine_similarity(self) -> None:
+        self.cosine_similarity = torch.nn.functional.cosine_similarity(self.previous_residual_subsampled, self.current_residual_subsampled, dim=-1).mean().item()
+
+    def derivative_approximation(self, sb_holder: SortblockHolder, output_raw: Union[torch.Tensor, tuple[torch.Tensor, ...]], orig_inputs: Union[torch.Tensor, tuple[torch.Tensor, ...]]):
+        activation_distance = sb_holder.activated_steps[-1] - sb_holder.activated_steps[-2]
+        # make tuple if not already tuple, so that works with both single and double blocks
+        if not isinstance(output_raw, tuple):
+            output_raw = (output_raw,)
+        if not isinstance(orig_inputs, tuple):
+            orig_inputs = (orig_inputs,)
+
+        for i, (output, x) in enumerate(zip(output_raw, orig_inputs)):
+            feature = output.clone() - x
+            has_previous_taylor_factor = self.previous_taylor_factors.get(i, None) is not None
+            # NOTE: not sure why - 2, but that's what's in the original implementation. Maybe consider changing values?
+            if has_previous_taylor_factor and sb_holder.step_count > (sb_holder.first_enhance - 2):
+                self.current_taylor_factors[i] = (
+                    feature - self.previous_taylor_factors[i]
+                ) / activation_distance
+
+            self.previous_taylor_factors[i] = feature
+
+    def use_taylor_formula(self, sb_holder: SortblockHolder) -> tuple[torch.Tensor, ...]:
+        step_distance = sb_holder.step_count - sb_holder.activated_steps[-1]
+
+        output_predicted = []
+
+        for key in self.previous_taylor_factors.keys():
+            previous_tf = self.previous_taylor_factors[key]
+            current_tf = self.current_taylor_factors[key]
+            predicted = taylor_formula(previous_tf, 0, step_distance)
+            predicted += taylor_formula(current_tf, 1, step_distance)
+            output_predicted.append(predicted)
+
+        return tuple(output_predicted)
+
+    def reset(self):
+        self.recompute = False
+        self.current_residual_subsampled = None
+        self.previous_residual_subsampled = None
+        self.cosine_similarity = None
+        self.previous_taylor_factors = {}
+        self.current_taylor_factors = {}
+
+def taylor_formula(taylor_factor: torch.Tensor, i: int, step_distance: int):
+    return (
+        (1 / math.factorial(i))
+        * taylor_factor
+        * (step_distance ** i)
+    )
+
+class SortblockHolder:
+    def __init__(self, start_predict_ratio: float, end_predict_ratio: float, policy_refresh_interval: int,
+                 start_percent: float, end_percent: float, subsample_factor: int=8, verbose: bool=False):
+        self.start_predict_ratio = start_predict_ratio
+        self.end_predict_ratio = end_predict_ratio
+        self.start_percent = start_percent
+        self.end_percent = end_percent
+        self.subsample_factor = subsample_factor
+        self.verbose = verbose
+
+        # NOTE: number represents steps
+        self.policy_refresh_interval = policy_refresh_interval
+        self.active_policy_refresh_interval = 1
+        self.first_enhance = 3  # NOTE: this value is 2 higher than the one actually used in code (subtracted by 2 in derivative_approximation)
+        # timestep values
+        self.start_t = 0.0
+        self.end_t = 0.0
+        self.curr_t = 0.0
+        # control values
+        self.initial_step = True
+        self.step_count = 0
+        self.activated_steps: list[int] = [0]
+        self.step_modulus = 0
+        self.do_sortblock = False
+        self.active_steps = 0
+        self.predict_ratios = []
+        self.predict_start_index = 0
+
+        # cache values
+        self.all_blocks = []
+        self.blocks_per_type = {}
+
+    def add_to_all_blocks(self, block):
+        self.all_blocks.append(block)
+
+    def add_to_blocks_per_type(self, block, block_type: str):
+        self.blocks_per_type.setdefault(block_type, []).append(block)
+
+    def prepare_timesteps(self, model_sampling):
+        self.start_t = model_sampling.percent_to_sigma(self.start_percent)
+        self.end_t = model_sampling.percent_to_sigma(self.end_percent)
+        return self
+
+    def check_if_within_timesteps(self, timestep: Union[float, torch.Tensor]) -> bool:
+        return (timestep <= self.start_t).item() and (timestep > self.end_t).item()
+
+    def update_should_do_sortblock(self, timestep: float) -> bool:
+        self.do_sortblock = (timestep[0] <= self.start_t).item() and (timestep[0] > self.end_t).item()
+        self.curr_t = timestep
+        if self.do_sortblock:
+            self.active_policy_refresh_interval = self.policy_refresh_interval
+        else:
+            self.active_policy_refresh_interval = 1
+        self.update_step_modulus()
+        return self.do_sortblock
+
+    def update_step_modulus(self):
+        self.step_modulus = int(self.step_count % self.active_policy_refresh_interval)
+
+    def should_do_sortblock(self) -> bool:
+        return self.do_sortblock
+
+    def reset(self):
+        self.initial_step = True
+        self.curr_t = 0.0
+        logging.info(f"Sortblock: resetting {len(self.all_blocks)} blocks")
+        for block in self.all_blocks:
+            clean_block(block)
+        self.all_blocks = []
+        self.blocks_per_type = {}
+        self.step_count = 0
+        self.activated_steps = [0]
+        self.step_modulus = 0
+        self.active_steps = 0
+        self.predict_ratios = []
+        self.do_sortblock = False
+        self.predict_start_index = 0
+        return self
+
+    def clone(self):
+        return SortblockHolder(start_predict_ratio=self.start_predict_ratio, end_predict_ratio=self.end_predict_ratio, policy_refresh_interval=self.policy_refresh_interval,
+            start_percent=self.start_percent, end_percent=self.end_percent, subsample_factor=self.subsample_factor,
+            verbose=self.verbose)
+
+
+class SortblockNode(io.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="Sortblock",
+            display_name="Sortblock",
+            description="A homebrew version of EasyCache - even 'easier' version of EasyCache to implement. Overall works worse than EasyCache, but better in some rare cases AND universal compatibility with everything in ComfyUI.",
+            category="advanced/debug/model",
+            is_experimental=True,
+            inputs=[
+                io.Model.Input("model", tooltip="The model to add Sortblock to."),
+                io.Float.Input("predict_ratio", min=0.0, default=0.8, max=3.0, step=0.01, tooltip="The ratio of blocks to predict."),
+                io.Int.Input("policy_refresh_interval", min=3, default=5, max=100, step=1, tooltip="The interval at which to refresh the policy."),
+                io.Float.Input("start_percent", min=0.0, default=0.15, max=1.0, step=0.01, tooltip="The relative sampling step to begin use of Sortblock."),
+                io.Float.Input("end_percent", min=0.0, default=0.95, max=1.0, step=0.01, tooltip="The relative sampling step to end use of Sortblock."),
+                io.Boolean.Input("verbose", default=False, tooltip="Whether to log verbose information."),
+            ],
+            outputs=[
+                io.Model.Output(tooltip="The model with Sortblock."),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, predict_ratio: float, policy_refresh_interval: int, start_percent: float, end_percent: float, verbose: bool) -> io.NodeOutput:
+        # TODO: check for specific flavors of supported models
+        model = model.clone()
+        model.model_options["transformer_options"]["sortblock"] = SortblockHolder(start_predict_ratio=predict_ratio, end_predict_ratio=predict_ratio, policy_refresh_interval=policy_refresh_interval,
+                                                                                  start_percent=start_percent, end_percent=end_percent, subsample_factor=8, verbose=verbose)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.PREDICT_NOISE, "sortblock", prepare_noise_wrapper)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, "sortblock", outer_sample_wrapper)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, "sortblock", model_forward_wrapper)
+        return io.NodeOutput(model)
+
+
+class SortblockScaledNode(io.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="SortblockScaled",
+            display_name="SortblockScaled",
+            description="A homebrew version of EasyCache - even 'easier' version of EasyCache to implement. Overall works worse than EasyCache, but better in some rare cases AND universal compatibility with everything in ComfyUI.",
+            category="advanced/debug/model",
+            is_experimental=True,
+            inputs=[
+                io.Model.Input("model", tooltip="The model to add Sortblock to."),
+                io.Float.Input("start_predict_ratio", min=0.0, default=0.2, max=1.0, step=0.01, tooltip="The ratio of blocks to predict."),
+                io.Float.Input("end_predict_ratio", min=0.0, default=0.9, max=1.0, step=0.01, tooltip="The ratio of blocks to predict."),
+                io.Int.Input("policy_refresh_interval", min=3, default=5, max=100, step=1, tooltip="The interval at which to refresh the policy."),
+                io.Float.Input("start_percent", min=0.0, default=0.15, max=1.0, step=0.01, tooltip="The relative sampling step to begin use of Sortblock."),
+                io.Float.Input("end_percent", min=0.0, default=0.95, max=1.0, step=0.01, tooltip="The relative sampling step to end use of Sortblock."),
+                io.Boolean.Input("verbose", default=False, tooltip="Whether to log verbose information."),
+            ],
+            outputs=[
+                io.Model.Output(tooltip="The model with Sortblock."),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, start_predict_ratio: float, end_predict_ratio: float, policy_refresh_interval: int, start_percent: float, end_percent: float, verbose: bool) -> io.NodeOutput:
+        # TODO: check for specific flavors of supported models
+        model = model.clone()
+        model.model_options["transformer_options"]["sortblock"] = SortblockHolder(start_predict_ratio, end_predict_ratio, policy_refresh_interval, start_percent, end_percent, subsample_factor=8, verbose=verbose)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.PREDICT_NOISE, "sortblock", prepare_noise_wrapper)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, "sortblock", outer_sample_wrapper)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, "sortblock", model_forward_wrapper)
+        return io.NodeOutput(model)
+
+
+class SortblockExtension(ComfyExtension):
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            SortblockNode,
+            SortblockScaledNode,
+        ]
+
+def comfy_entrypoint():
+    return SortblockExtension()

comfy_extras/nodes_string.py

@@ -1,77 +1,91 @@
 import re
+from typing_extensions import override
 
-from comfy.comfy_types.node_typing import IO
+from comfy_api.latest import ComfyExtension, io
 
-class StringConcatenate():
+
+class StringConcatenate(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string_a": (IO.STRING, {"multiline": True}),
-                "string_b": (IO.STRING, {"multiline": True}),
-                "delimiter": (IO.STRING, {"multiline": False, "default": ""})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringConcatenate",
+            display_name="Concatenate",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string_a", multiline=True),
+                io.String.Input("string_b", multiline=True),
+                io.String.Input("delimiter", multiline=False, default=""),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string_a, string_b, delimiter, **kwargs):
-        return delimiter.join((string_a, string_b)),
-
-class StringSubstring():
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "start": (IO.INT, {}),
-                "end": (IO.INT, {}),
-            }
-        }
+    def execute(cls, string_a, string_b, delimiter):
+        return io.NodeOutput(delimiter.join((string_a, string_b)))
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
 
-    def execute(self, string, start, end, **kwargs):
-        return string[start:end],
-
-class StringLength():
+class StringSubstring(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringSubstring",
+            display_name="Substring",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.Int.Input("start"),
+                io.Int.Input("end"),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.INT,)
-    RETURN_NAMES = ("length",)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string, **kwargs):
-        length = len(string)
-
-        return length,
-
-class CaseConverter():
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "mode": (IO.COMBO, {"options": ["UPPERCASE", "lowercase", "Capitalize", "Title Case"]})
-            }
-        }
+    def execute(cls, string, start, end):
+        return io.NodeOutput(string[start:end])
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
 
-    def execute(self, string, mode, **kwargs):
+class StringLength(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringLength",
+            display_name="Length",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+            ],
+            outputs=[
+                io.Int.Output(display_name="length"),
+            ]
+        )
+
+    @classmethod
+    def execute(cls, string):
+        return io.NodeOutput(len(string))
+
+
+class CaseConverter(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="CaseConverter",
+            display_name="Case Converter",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.Combo.Input("mode", options=["UPPERCASE", "lowercase", "Capitalize", "Title Case"]),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
+
+    @classmethod
+    def execute(cls, string, mode):
         if mode == "UPPERCASE":
             result = string.upper()
         elif mode == "lowercase":
@@ -83,24 +97,27 @@ class CaseConverter():
         else:
             result = string
 
-        return result,
+        return io.NodeOutput(result)
 
 
-class StringTrim():
+class StringTrim(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "mode": (IO.COMBO, {"options": ["Both", "Left", "Right"]})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringTrim",
+            display_name="Trim",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.Combo.Input("mode", options=["Both", "Left", "Right"]),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string, mode, **kwargs):
+    @classmethod
+    def execute(cls, string, mode):
         if mode == "Both":
             result = string.strip()
         elif mode == "Left":
@@ -110,70 +127,78 @@ class StringTrim():
         else:
             result = string
 
-        return result,
+        return io.NodeOutput(result)
 
-class StringReplace():
+
+class StringReplace(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "find": (IO.STRING, {"multiline": True}),
-                "replace": (IO.STRING, {"multiline": True})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringReplace",
+            display_name="Replace",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.String.Input("find", multiline=True),
+                io.String.Input("replace", multiline=True),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string, find, replace, **kwargs):
-        result = string.replace(find, replace)
-        return result,
-
-
-class StringContains():
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "substring": (IO.STRING, {"multiline": True}),
-                "case_sensitive": (IO.BOOLEAN, {"default": True})
-            }
-        }
+    def execute(cls, string, find, replace):
+        return io.NodeOutput(string.replace(find, replace))
 
-    RETURN_TYPES = (IO.BOOLEAN,)
-    RETURN_NAMES = ("contains",)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
 
-    def execute(self, string, substring, case_sensitive, **kwargs):
+class StringContains(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringContains",
+            display_name="Contains",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.String.Input("substring", multiline=True),
+                io.Boolean.Input("case_sensitive", default=True),
+            ],
+            outputs=[
+                io.Boolean.Output(display_name="contains"),
+            ]
+        )
+
+    @classmethod
+    def execute(cls, string, substring, case_sensitive):
         if case_sensitive:
             contains = substring in string
         else:
             contains = substring.lower() in string.lower()
 
-        return contains,
+        return io.NodeOutput(contains)
 
 
-class StringCompare():
+class StringCompare(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string_a": (IO.STRING, {"multiline": True}),
-                "string_b": (IO.STRING, {"multiline": True}),
-                "mode": (IO.COMBO, {"options": ["Starts With", "Ends With", "Equal"]}),
-                "case_sensitive": (IO.BOOLEAN, {"default": True})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringCompare",
+            display_name="Compare",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string_a", multiline=True),
+                io.String.Input("string_b", multiline=True),
+                io.Combo.Input("mode", options=["Starts With", "Ends With", "Equal"]),
+                io.Boolean.Input("case_sensitive", default=True),
+            ],
+            outputs=[
+                io.Boolean.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.BOOLEAN,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string_a, string_b, mode, case_sensitive, **kwargs):
+    @classmethod
+    def execute(cls, string_a, string_b, mode, case_sensitive):
         if case_sensitive:
             a = string_a
             b = string_b
@@ -182,31 +207,34 @@ class StringCompare():
             b = string_b.lower()
 
         if mode == "Equal":
-            return a == b,
+            return io.NodeOutput(a == b)
         elif mode == "Starts With":
-            return a.startswith(b),
+            return io.NodeOutput(a.startswith(b))
         elif mode == "Ends With":
-            return a.endswith(b),
+            return io.NodeOutput(a.endswith(b))
 
-class RegexMatch():
+
+class RegexMatch(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "regex_pattern": (IO.STRING, {"multiline": True}),
-                "case_insensitive": (IO.BOOLEAN, {"default": True}),
-                "multiline": (IO.BOOLEAN, {"default": False}),
-                "dotall": (IO.BOOLEAN, {"default": False})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="RegexMatch",
+            display_name="Regex Match",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.String.Input("regex_pattern", multiline=True),
+                io.Boolean.Input("case_insensitive", default=True),
+                io.Boolean.Input("multiline", default=False),
+                io.Boolean.Input("dotall", default=False),
+            ],
+            outputs=[
+                io.Boolean.Output(display_name="matches"),
+            ]
+        )
 
-    RETURN_TYPES = (IO.BOOLEAN,)
-    RETURN_NAMES = ("matches",)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string, regex_pattern, case_insensitive, multiline, dotall, **kwargs):
+    @classmethod
+    def execute(cls, string, regex_pattern, case_insensitive, multiline, dotall):
         flags = 0
 
         if case_insensitive:
@@ -223,29 +251,32 @@ class RegexMatch():
         except re.error:
             result = False
 
-        return result,
+        return io.NodeOutput(result)
 
 
-class RegexExtract():
+class RegexExtract(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "regex_pattern": (IO.STRING, {"multiline": True}),
-                "mode": (IO.COMBO, {"options": ["First Match", "All Matches", "First Group", "All Groups"]}),
-                "case_insensitive": (IO.BOOLEAN, {"default": True}),
-                "multiline": (IO.BOOLEAN, {"default": False}),
-                "dotall": (IO.BOOLEAN, {"default": False}),
-                "group_index": (IO.INT, {"default": 1, "min": 0, "max": 100})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="RegexExtract",
+            display_name="Regex Extract",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.String.Input("regex_pattern", multiline=True),
+                io.Combo.Input("mode", options=["First Match", "All Matches", "First Group", "All Groups"]),
+                io.Boolean.Input("case_insensitive", default=True),
+                io.Boolean.Input("multiline", default=False),
+                io.Boolean.Input("dotall", default=False),
+                io.Int.Input("group_index", default=1, min=0, max=100),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string, regex_pattern, mode, case_insensitive, multiline, dotall, group_index, **kwargs):
+    @classmethod
+    def execute(cls, string, regex_pattern, mode, case_insensitive, multiline, dotall, group_index):
         join_delimiter = "\n"
 
         flags = 0
@@ -294,32 +325,33 @@ class RegexExtract():
         except re.error:
             result = ""
 
-        return result,
+        return io.NodeOutput(result)
 
 
-class RegexReplace():
-    DESCRIPTION = "Find and replace text using regex patterns."
+class RegexReplace(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "regex_pattern": (IO.STRING, {"multiline": True}),
-                "replace": (IO.STRING, {"multiline": True}),
-            },
-            "optional": {
-                "case_insensitive": (IO.BOOLEAN, {"default": True}),
-                "multiline": (IO.BOOLEAN, {"default": False}),
-                "dotall": (IO.BOOLEAN, {"default": False, "tooltip": "When enabled, the dot (.) character will match any character including newline characters. When disabled, dots won't match newlines."}),
-                "count": (IO.INT, {"default": 0, "min": 0, "max": 100, "tooltip": "Maximum number of replacements to make. Set to 0 to replace all occurrences (default). Set to 1 to replace only the first match, 2 for the first two matches, etc."}),
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="RegexReplace",
+            display_name="Regex Replace",
+            category="utils/string",
+            description="Find and replace text using regex patterns.",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.String.Input("regex_pattern", multiline=True),
+                io.String.Input("replace", multiline=True),
+                io.Boolean.Input("case_insensitive", default=True, optional=True),
+                io.Boolean.Input("multiline", default=False, optional=True),
+                io.Boolean.Input("dotall", default=False, optional=True, tooltip="When enabled, the dot (.) character will match any character including newline characters. When disabled, dots won't match newlines."),
+                io.Int.Input("count", default=0, min=0, max=100, optional=True, tooltip="Maximum number of replacements to make. Set to 0 to replace all occurrences (default). Set to 1 to replace only the first match, 2 for the first two matches, etc."),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string, regex_pattern, replace, case_insensitive=True, multiline=False, dotall=False, count=0, **kwargs):
+    @classmethod
+    def execute(cls, string, regex_pattern, replace, case_insensitive=True, multiline=False, dotall=False, count=0):
         flags = 0
 
         if case_insensitive:
@@ -329,32 +361,25 @@ class RegexReplace():
         if dotall:
             flags |= re.DOTALL
         result = re.sub(regex_pattern, replace, string, count=count, flags=flags)
-        return result,
+        return io.NodeOutput(result)
 
-NODE_CLASS_MAPPINGS = {
-    "StringConcatenate": StringConcatenate,
-    "StringSubstring": StringSubstring,
-    "StringLength": StringLength,
-    "CaseConverter": CaseConverter,
-    "StringTrim": StringTrim,
-    "StringReplace": StringReplace,
-    "StringContains": StringContains,
-    "StringCompare": StringCompare,
-    "RegexMatch": RegexMatch,
-    "RegexExtract": RegexExtract,
-    "RegexReplace": RegexReplace,
-}
 
-NODE_DISPLAY_NAME_MAPPINGS = {
-    "StringConcatenate": "Concatenate",
-    "StringSubstring": "Substring",
-    "StringLength": "Length",
-    "CaseConverter": "Case Converter",
-    "StringTrim": "Trim",
-    "StringReplace": "Replace",
-    "StringContains": "Contains",
-    "StringCompare": "Compare",
-    "RegexMatch": "Regex Match",
-    "RegexExtract": "Regex Extract",
-    "RegexReplace": "Regex Replace",
-}
+class StringExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            StringConcatenate,
+            StringSubstring,
+            StringLength,
+            CaseConverter,
+            StringTrim,
+            StringReplace,
+            StringContains,
+            StringCompare,
+            RegexMatch,
+            RegexExtract,
+            RegexReplace,
+        ]
+
+async def comfy_entrypoint() -> StringExtension:
+    return StringExtension()

comfy_extras/nodes_wan.py

@@ -9,29 +9,35 @@ import comfy.clip_vision
 import json
 import numpy as np
 from typing import Tuple
+from typing_extensions import override
+from comfy_api.latest import ComfyExtension, io
 
-class WanImageToVideo:
+class WanImageToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ),
-                             "start_image": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanImageToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+                io.Image.Input("start_image", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None) -> io.NodeOutput:
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         if start_image is not None:
             start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
@@ -51,32 +57,36 @@ class WanImageToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent)
+        return io.NodeOutput(positive, negative, out_latent)
 
 
-class WanFunControlToVideo:
+class WanFunControlToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ),
-                             "start_image": ("IMAGE", ),
-                             "control_video": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanFunControlToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+                io.Image.Input("start_image", optional=True),
+                io.Image.Input("control_video", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None, control_video=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None, control_video=None) -> io.NodeOutput:
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         concat_latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         concat_latent = comfy.latent_formats.Wan21().process_out(concat_latent)
@@ -101,33 +111,103 @@ class WanFunControlToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent)
+        return io.NodeOutput(positive, negative, out_latent)
 
-class WanFirstLastFrameToVideo:
+class Wan22FunControlToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"clip_vision_start_image": ("CLIP_VISION_OUTPUT", ),
-                             "clip_vision_end_image": ("CLIP_VISION_OUTPUT", ),
-                             "start_image": ("IMAGE", ),
-                             "end_image": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="Wan22FunControlToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Image.Input("ref_image", optional=True),
+                io.Image.Input("control_video", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, ref_image=None, start_image=None, control_video=None) -> io.NodeOutput:
+        spacial_scale = vae.spacial_compression_encode()
+        latent_channels = vae.latent_channels
+        latent = torch.zeros([batch_size, latent_channels, ((length - 1) // 4) + 1, height // spacial_scale, width // spacial_scale], device=comfy.model_management.intermediate_device())
+        concat_latent = torch.zeros([batch_size, latent_channels, ((length - 1) // 4) + 1, height // spacial_scale, width // spacial_scale], device=comfy.model_management.intermediate_device())
+        if latent_channels == 48:
+            concat_latent = comfy.latent_formats.Wan22().process_out(concat_latent)
+        else:
+            concat_latent = comfy.latent_formats.Wan21().process_out(concat_latent)
+        concat_latent = concat_latent.repeat(1, 2, 1, 1, 1)
+        mask = torch.ones((1, 1, latent.shape[2] * 4, latent.shape[-2], latent.shape[-1]))
 
-    CATEGORY = "conditioning/video_models"
+        if start_image is not None:
+            start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            concat_latent_image = vae.encode(start_image[:, :, :, :3])
+            concat_latent[:,latent_channels:,:concat_latent_image.shape[2]] = concat_latent_image[:,:,:concat_latent.shape[2]]
+            mask[:, :, :start_image.shape[0] + 3] = 0.0
 
-    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, end_image=None, clip_vision_start_image=None, clip_vision_end_image=None):
-        latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+        ref_latent = None
+        if ref_image is not None:
+            ref_image = comfy.utils.common_upscale(ref_image[:1].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            ref_latent = vae.encode(ref_image[:, :, :, :3])
+
+        if control_video is not None:
+            control_video = comfy.utils.common_upscale(control_video[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            concat_latent_image = vae.encode(control_video[:, :, :, :3])
+            concat_latent[:,:latent_channels,:concat_latent_image.shape[2]] = concat_latent_image[:,:,:concat_latent.shape[2]]
+
+        mask = mask.view(1, mask.shape[2] // 4, 4, mask.shape[3], mask.shape[4]).transpose(1, 2)
+        positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent, "concat_mask": mask, "concat_mask_index": latent_channels})
+        negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent, "concat_mask": mask, "concat_mask_index": latent_channels})
+
+        if ref_latent is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"reference_latents": [ref_latent]}, append=True)
+            negative = node_helpers.conditioning_set_values(negative, {"reference_latents": [ref_latent]}, append=True)
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return io.NodeOutput(positive, negative, out_latent)
+
+class WanFirstLastFrameToVideo(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanFirstLastFrameToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_start_image", optional=True),
+                io.ClipVisionOutput.Input("clip_vision_end_image", optional=True),
+                io.Image.Input("start_image", optional=True),
+                io.Image.Input("end_image", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, end_image=None, clip_vision_start_image=None, clip_vision_end_image=None) -> io.NodeOutput:
+        spacial_scale = vae.spacial_compression_encode()
+        latent = torch.zeros([batch_size, vae.latent_channels, ((length - 1) // 4) + 1, height // spacial_scale, width // spacial_scale], device=comfy.model_management.intermediate_device())
         if start_image is not None:
             start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
         if end_image is not None:
@@ -167,62 +247,70 @@ class WanFirstLastFrameToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent)
+        return io.NodeOutput(positive, negative, out_latent)
 
 
-class WanFunInpaintToVideo:
+class WanFunInpaintToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ),
-                             "start_image": ("IMAGE", ),
-                             "end_image": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanFunInpaintToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+                io.Image.Input("start_image", optional=True),
+                io.Image.Input("end_image", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, end_image=None, clip_vision_output=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, end_image=None, clip_vision_output=None) -> io.NodeOutput:
         flfv = WanFirstLastFrameToVideo()
-        return flfv.encode(positive, negative, vae, width, height, length, batch_size, start_image=start_image, end_image=end_image, clip_vision_start_image=clip_vision_output)
+        return flfv.execute(positive, negative, vae, width, height, length, batch_size, start_image=start_image, end_image=end_image, clip_vision_start_image=clip_vision_output)
 
 
-class WanVaceToVideo:
+class WanVaceToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                             "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1000.0, "step": 0.01}),
-                },
-                "optional": {"control_video": ("IMAGE", ),
-                             "control_masks": ("MASK", ),
-                             "reference_image": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanVaceToVideo",
+            category="conditioning/video_models",
+            is_experimental=True,
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Float.Input("strength", default=1.0, min=0.0, max=1000.0, step=0.01),
+                io.Image.Input("control_video", optional=True),
+                io.Mask.Input("control_masks", optional=True),
+                io.Image.Input("reference_image", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+                io.Int.Output(display_name="trim_latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT", "INT")
-    RETURN_NAMES = ("positive", "negative", "latent", "trim_latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    EXPERIMENTAL = True
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, strength, control_video=None, control_masks=None, reference_image=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, strength, control_video=None, control_masks=None, reference_image=None) -> io.NodeOutput:
         latent_length = ((length - 1) // 4) + 1
         if control_video is not None:
             control_video = comfy.utils.common_upscale(control_video[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
@@ -279,52 +367,59 @@ class WanVaceToVideo:
         latent = torch.zeros([batch_size, 16, latent_length, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent, trim_latent)
+        return io.NodeOutput(positive, negative, out_latent, trim_latent)
 
-class TrimVideoLatent:
+class TrimVideoLatent(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "samples": ("LATENT",),
-                              "trim_amount": ("INT", {"default": 0, "min": 0, "max": 99999}),
-                             }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TrimVideoLatent",
+            category="latent/video",
+            is_experimental=True,
+            inputs=[
+                io.Latent.Input("samples"),
+                io.Int.Input("trim_amount", default=0, min=0, max=99999),
+            ],
+            outputs=[
+                io.Latent.Output(),
+            ],
+        )
 
-    RETURN_TYPES = ("LATENT",)
-    FUNCTION = "op"
-
-    CATEGORY = "latent/video"
-
-    EXPERIMENTAL = True
-
-    def op(self, samples, trim_amount):
+    @classmethod
+    def execute(cls, samples, trim_amount) -> io.NodeOutput:
         samples_out = samples.copy()
 
         s1 = samples["samples"]
         samples_out["samples"] = s1[:, :, trim_amount:]
-        return (samples_out,)
+        return io.NodeOutput(samples_out)
 
-class WanCameraImageToVideo:
+class WanCameraImageToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ),
-                             "start_image": ("IMAGE", ),
-                             "camera_conditions": ("WAN_CAMERA_EMBEDDING", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanCameraImageToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+                io.Image.Input("start_image", optional=True),
+                io.WanCameraEmbedding.Input("camera_conditions", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None, camera_conditions=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None, camera_conditions=None) -> io.NodeOutput:
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         concat_latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         concat_latent = comfy.latent_formats.Wan21().process_out(concat_latent)
@@ -333,9 +428,12 @@ class WanCameraImageToVideo:
             start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
             concat_latent_image = vae.encode(start_image[:, :, :, :3])
             concat_latent[:,:,:concat_latent_image.shape[2]] = concat_latent_image[:,:,:concat_latent.shape[2]]
+            mask = torch.ones((1, 1, latent.shape[2] * 4, latent.shape[-2], latent.shape[-1]))
+            mask[:, :, :start_image.shape[0] + 3] = 0.0
+            mask = mask.view(1, mask.shape[2] // 4, 4, mask.shape[3], mask.shape[4]).transpose(1, 2)
 
-            positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent})
-            negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent})
+            positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent, "concat_mask": mask})
+            negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent, "concat_mask": mask})
 
         if camera_conditions is not None:
             positive = node_helpers.conditioning_set_values(positive, {'camera_conditions': camera_conditions})
@@ -347,29 +445,34 @@ class WanCameraImageToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent)
+        return io.NodeOutput(positive, negative, out_latent)
 
-class WanPhantomSubjectToVideo:
+class WanPhantomSubjectToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"images": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanPhantomSubjectToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Image.Input("images", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative_text"),
+                io.Conditioning.Output(display_name="negative_img_text"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative_text", "negative_img_text", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, images):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, images) -> io.NodeOutput:
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         cond2 = negative
         if images is not None:
@@ -385,7 +488,7 @@ class WanPhantomSubjectToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, cond2, negative, out_latent)
+        return io.NodeOutput(positive, cond2, negative, out_latent)
 
 def parse_json_tracks(tracks):
     """Parse JSON track data into a standardized format"""
@@ -598,39 +701,41 @@ def patch_motion(
 
     return out_mask_full, out_feature_full
 
-class WanTrackToVideo:
+class WanTrackToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {
-                    "positive": ("CONDITIONING", ),
-                    "negative": ("CONDITIONING", ),
-                    "vae": ("VAE", ),
-                    "tracks": ("STRING", {"multiline": True, "default": "[]"}),
-                    "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                    "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                    "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                    "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                    "temperature": ("FLOAT", {"default": 220.0, "min": 1.0, "max": 1000.0, "step": 0.1}),
-                    "topk": ("INT", {"default": 2, "min": 1, "max": 10}),
-                    "start_image": ("IMAGE", ),
-                },
-                "optional": {
-                    "clip_vision_output": ("CLIP_VISION_OUTPUT", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanTrackToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.String.Input("tracks", multiline=True, default="[]"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Float.Input("temperature", default=220.0, min=1.0, max=1000.0, step=0.1),
+                io.Int.Input("topk", default=2, min=1, max=10),
+                io.Image.Input("start_image"),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, tracks, width, height, length, batch_size,
-               temperature, topk, start_image=None, clip_vision_output=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, tracks, width, height, length, batch_size,
+               temperature, topk, start_image=None, clip_vision_output=None) -> io.NodeOutput:
 
         tracks_data = parse_json_tracks(tracks)
 
         if not tracks_data:
-            return WanImageToVideo().encode(positive, negative, vae, width, height, length, batch_size, start_image=start_image, clip_vision_output=clip_vision_output)
+            return WanImageToVideo().execute(positive, negative, vae, width, height, length, batch_size, start_image=start_image, clip_vision_output=clip_vision_output)
 
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8],
                            device=comfy.model_management.intermediate_device())
@@ -684,34 +789,259 @@ class WanTrackToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent)
+        return io.NodeOutput(positive, negative, out_latent)
 
 
-class Wan22ImageToVideoLatent:
+def linear_interpolation(features, input_fps, output_fps, output_len=None):
+    """
+    features: shape=[1, T, 512]
+    input_fps: fps for audio, f_a
+    output_fps: fps for video, f_m
+    output_len: video length
+    """
+    features = features.transpose(1, 2)  # [1, 512, T]
+    seq_len = features.shape[2] / float(input_fps)  # T/f_a
+    if output_len is None:
+        output_len = int(seq_len * output_fps)  # f_m*T/f_a
+    output_features = torch.nn.functional.interpolate(
+        features, size=output_len, align_corners=True,
+        mode='linear')  # [1, 512, output_len]
+    return output_features.transpose(1, 2)  # [1, output_len, 512]
+
+
+def get_sample_indices(original_fps,
+                       total_frames,
+                       target_fps,
+                       num_sample,
+                       fixed_start=None):
+    required_duration = num_sample / target_fps
+    required_origin_frames = int(np.ceil(required_duration * original_fps))
+    if required_duration > total_frames / original_fps:
+        raise ValueError("required_duration must be less than video length")
+
+    if not fixed_start is None and fixed_start >= 0:
+        start_frame = fixed_start
+    else:
+        max_start = total_frames - required_origin_frames
+        if max_start < 0:
+            raise ValueError("video length is too short")
+        start_frame = np.random.randint(0, max_start + 1)
+    start_time = start_frame / original_fps
+
+    end_time = start_time + required_duration
+    time_points = np.linspace(start_time, end_time, num_sample, endpoint=False)
+
+    frame_indices = np.round(np.array(time_points) * original_fps).astype(int)
+    frame_indices = np.clip(frame_indices, 0, total_frames - 1)
+    return frame_indices
+
+
+def get_audio_embed_bucket_fps(audio_embed, fps=16, batch_frames=81, m=0, video_rate=30):
+    num_layers, audio_frame_num, audio_dim = audio_embed.shape
+
+    if num_layers > 1:
+        return_all_layers = True
+    else:
+        return_all_layers = False
+
+    scale = video_rate / fps
+
+    min_batch_num = int(audio_frame_num / (batch_frames * scale)) + 1
+
+    bucket_num = min_batch_num * batch_frames
+    padd_audio_num = math.ceil(min_batch_num * batch_frames / fps * video_rate) - audio_frame_num
+    batch_idx = get_sample_indices(
+        original_fps=video_rate,
+        total_frames=audio_frame_num + padd_audio_num,
+        target_fps=fps,
+        num_sample=bucket_num,
+        fixed_start=0)
+    batch_audio_eb = []
+    audio_sample_stride = int(video_rate / fps)
+    for bi in batch_idx:
+        if bi < audio_frame_num:
+
+            chosen_idx = list(
+                range(bi - m * audio_sample_stride, bi + (m + 1) * audio_sample_stride, audio_sample_stride))
+            chosen_idx = [0 if c < 0 else c for c in chosen_idx]
+            chosen_idx = [
+                audio_frame_num - 1 if c >= audio_frame_num else c
+                for c in chosen_idx
+            ]
+
+            if return_all_layers:
+                frame_audio_embed = audio_embed[:, chosen_idx].flatten(
+                    start_dim=-2, end_dim=-1)
+            else:
+                frame_audio_embed = audio_embed[0][chosen_idx].flatten()
+        else:
+            frame_audio_embed = torch.zeros([audio_dim * (2 * m + 1)], device=audio_embed.device) if not return_all_layers \
+                else torch.zeros([num_layers, audio_dim * (2 * m + 1)], device=audio_embed.device)
+        batch_audio_eb.append(frame_audio_embed)
+    batch_audio_eb = torch.cat([c.unsqueeze(0) for c in batch_audio_eb], dim=0)
+
+    return batch_audio_eb, min_batch_num
+
+
+def wan_sound_to_video(positive, negative, vae, width, height, length, batch_size, frame_offset=0, ref_image=None, audio_encoder_output=None, control_video=None, ref_motion=None, ref_motion_latent=None):
+    latent_t = ((length - 1) // 4) + 1
+    if audio_encoder_output is not None:
+        feat = torch.cat(audio_encoder_output["encoded_audio_all_layers"])
+        video_rate = 30
+        fps = 16
+        feat = linear_interpolation(feat, input_fps=50, output_fps=video_rate)
+        batch_frames = latent_t * 4
+        audio_embed_bucket, num_repeat = get_audio_embed_bucket_fps(feat, fps=fps, batch_frames=batch_frames, m=0, video_rate=video_rate)
+        audio_embed_bucket = audio_embed_bucket.unsqueeze(0)
+        if len(audio_embed_bucket.shape) == 3:
+            audio_embed_bucket = audio_embed_bucket.permute(0, 2, 1)
+        elif len(audio_embed_bucket.shape) == 4:
+            audio_embed_bucket = audio_embed_bucket.permute(0, 2, 3, 1)
+
+        audio_embed_bucket = audio_embed_bucket[:, :, :, frame_offset:frame_offset + batch_frames]
+        if audio_embed_bucket.shape[3] > 0:
+            positive = node_helpers.conditioning_set_values(positive, {"audio_embed": audio_embed_bucket})
+            negative = node_helpers.conditioning_set_values(negative, {"audio_embed": audio_embed_bucket * 0.0})
+            frame_offset += batch_frames
+
+    if ref_image is not None:
+        ref_image = comfy.utils.common_upscale(ref_image[:1].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+        ref_latent = vae.encode(ref_image[:, :, :, :3])
+        positive = node_helpers.conditioning_set_values(positive, {"reference_latents": [ref_latent]}, append=True)
+        negative = node_helpers.conditioning_set_values(negative, {"reference_latents": [ref_latent]}, append=True)
+
+    if ref_motion is not None:
+        if ref_motion.shape[0] > 73:
+            ref_motion = ref_motion[-73:]
+
+        ref_motion = comfy.utils.common_upscale(ref_motion.movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+
+        if ref_motion.shape[0] < 73:
+            r = torch.ones([73, height, width, 3]) * 0.5
+            r[-ref_motion.shape[0]:] = ref_motion
+            ref_motion = r
+
+        ref_motion_latent = vae.encode(ref_motion[:, :, :, :3])
+
+    if ref_motion_latent is not None:
+        ref_motion_latent = ref_motion_latent[:, :, -19:]
+        positive = node_helpers.conditioning_set_values(positive, {"reference_motion": ref_motion_latent})
+        negative = node_helpers.conditioning_set_values(negative, {"reference_motion": ref_motion_latent})
+
+    latent = torch.zeros([batch_size, 16, latent_t, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+
+    control_video_out = comfy.latent_formats.Wan21().process_out(torch.zeros_like(latent))
+    if control_video is not None:
+        control_video = comfy.utils.common_upscale(control_video[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+        control_video = vae.encode(control_video[:, :, :, :3])
+        control_video_out[:, :, :control_video.shape[2]] = control_video
+
+    # TODO: check if zero is better than none if none provided
+    positive = node_helpers.conditioning_set_values(positive, {"control_video": control_video_out})
+    negative = node_helpers.conditioning_set_values(negative, {"control_video": control_video_out})
+
+    out_latent = {}
+    out_latent["samples"] = latent
+    return positive, negative, out_latent, frame_offset
+
+
+class WanSoundImageToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"vae": ("VAE", ),
-                             "width": ("INT", {"default": 1280, "min": 32, "max": nodes.MAX_RESOLUTION, "step": 32}),
-                             "height": ("INT", {"default": 704, "min": 32, "max": nodes.MAX_RESOLUTION, "step": 32}),
-                             "length": ("INT", {"default": 49, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"start_image": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanSoundImageToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=77, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.AudioEncoderOutput.Input("audio_encoder_output", optional=True),
+                io.Image.Input("ref_image", optional=True),
+                io.Image.Input("control_video", optional=True),
+                io.Image.Input("ref_motion", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+            is_experimental=True,
+        )
+
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, ref_image=None, audio_encoder_output=None, control_video=None, ref_motion=None) -> io.NodeOutput:
+        positive, negative, out_latent, frame_offset = wan_sound_to_video(positive, negative, vae, width, height, length, batch_size, ref_image=ref_image, audio_encoder_output=audio_encoder_output,
+                                                                          control_video=control_video, ref_motion=ref_motion)
+        return io.NodeOutput(positive, negative, out_latent)
 
 
-    RETURN_TYPES = ("LATENT",)
-    FUNCTION = "encode"
+class WanSoundImageToVideoExtend(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanSoundImageToVideoExtend",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("length", default=77, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Latent.Input("video_latent"),
+                io.AudioEncoderOutput.Input("audio_encoder_output", optional=True),
+                io.Image.Input("ref_image", optional=True),
+                io.Image.Input("control_video", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+            is_experimental=True,
+        )
 
-    CATEGORY = "conditioning/inpaint"
+    @classmethod
+    def execute(cls, positive, negative, vae, length, video_latent, ref_image=None, audio_encoder_output=None, control_video=None) -> io.NodeOutput:
+        video_latent = video_latent["samples"]
+        width = video_latent.shape[-1] * 8
+        height = video_latent.shape[-2] * 8
+        batch_size = video_latent.shape[0]
+        frame_offset = video_latent.shape[-3] * 4
+        positive, negative, out_latent, frame_offset = wan_sound_to_video(positive, negative, vae, width, height, length, batch_size, frame_offset=frame_offset, ref_image=ref_image, audio_encoder_output=audio_encoder_output,
+                                                                          control_video=control_video, ref_motion=None, ref_motion_latent=video_latent)
+        return io.NodeOutput(positive, negative, out_latent)
 
-    def encode(self, vae, width, height, length, batch_size, start_image=None):
+
+class Wan22ImageToVideoLatent(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="Wan22ImageToVideoLatent",
+            category="conditioning/inpaint",
+            inputs=[
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=1280, min=32, max=nodes.MAX_RESOLUTION, step=32),
+                io.Int.Input("height", default=704, min=32, max=nodes.MAX_RESOLUTION, step=32),
+                io.Int.Input("length", default=49, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Image.Input("start_image", optional=True),
+            ],
+            outputs=[
+                io.Latent.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, vae, width, height, length, batch_size, start_image=None) -> io.NodeOutput:
         latent = torch.zeros([1, 48, ((length - 1) // 4) + 1, height // 16, width // 16], device=comfy.model_management.intermediate_device())
 
         if start_image is None:
             out_latent = {}
             out_latent["samples"] = latent
-            return (out_latent,)
+            return io.NodeOutput(out_latent)
 
         mask = torch.ones([latent.shape[0], 1, ((length - 1) // 4) + 1, latent.shape[-2], latent.shape[-1]], device=comfy.model_management.intermediate_device())
 
@@ -726,18 +1056,27 @@ class Wan22ImageToVideoLatent:
         latent = latent_format.process_out(latent) * mask + latent * (1.0 - mask)
         out_latent["samples"] = latent.repeat((batch_size, ) + (1,) * (latent.ndim - 1))
         out_latent["noise_mask"] = mask.repeat((batch_size, ) + (1,) * (mask.ndim - 1))
-        return (out_latent,)
+        return io.NodeOutput(out_latent)
 
 
-NODE_CLASS_MAPPINGS = {
-    "WanTrackToVideo": WanTrackToVideo,
-    "WanImageToVideo": WanImageToVideo,
-    "WanFunControlToVideo": WanFunControlToVideo,
-    "WanFunInpaintToVideo": WanFunInpaintToVideo,
-    "WanFirstLastFrameToVideo": WanFirstLastFrameToVideo,
-    "WanVaceToVideo": WanVaceToVideo,
-    "TrimVideoLatent": TrimVideoLatent,
-    "WanCameraImageToVideo": WanCameraImageToVideo,
-    "WanPhantomSubjectToVideo": WanPhantomSubjectToVideo,
-    "Wan22ImageToVideoLatent": Wan22ImageToVideoLatent,
-}
+class WanExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            WanTrackToVideo,
+            WanImageToVideo,
+            WanFunControlToVideo,
+            Wan22FunControlToVideo,
+            WanFunInpaintToVideo,
+            WanFirstLastFrameToVideo,
+            WanVaceToVideo,
+            TrimVideoLatent,
+            WanCameraImageToVideo,
+            WanPhantomSubjectToVideo,
+            WanSoundImageToVideo,
+            WanSoundImageToVideoExtend,
+            Wan22ImageToVideoLatent,
+        ]
+
+async def comfy_entrypoint() -> WanExtension:
+    return WanExtension()

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.49"
+__version__ = "0.3.56"

folder_paths.py

@@ -46,6 +46,10 @@ folder_names_and_paths["photomaker"] = ([os.path.join(models_dir, "photomaker")]
 
 folder_names_and_paths["classifiers"] = ([os.path.join(models_dir, "classifiers")], {""})
 
+folder_names_and_paths["model_patches"] = ([os.path.join(models_dir, "model_patches")], supported_pt_extensions)
+
+folder_names_and_paths["audio_encoders"] = ([os.path.join(models_dir, "audio_encoders")], supported_pt_extensions)
+
 output_directory = os.path.join(base_path, "output")
 temp_directory = os.path.join(base_path, "temp")
 input_directory = os.path.join(base_path, "input")
@@ -275,7 +279,10 @@ def filter_files_extensions(files: Collection[str], extensions: Collection[str])
 
 
 
-def get_full_path(folder_name: str, filename: str, allow_missing: bool = False) -> str | None:
+def get_full_path(folder_name: str, filename: str) -> str | None:
+    """
+    Get the full path of a file in a folder, has to be a file
+    """
     global folder_names_and_paths
     folder_name = map_legacy(folder_name)
     if folder_name not in folder_names_and_paths:
@@ -288,8 +295,6 @@ def get_full_path(folder_name: str, filename: str, allow_missing: bool = False)
             return full_path
         elif os.path.islink(full_path):
             logging.warning("WARNING path {} exists but doesn't link anywhere, skipping.".format(full_path))
-        elif allow_missing:
-            return full_path
 
     return None
 
@@ -304,27 +309,6 @@ def get_full_path_or_raise(folder_name: str, filename: str) -> str:
     return full_path
 
 
-def get_relative_path(full_path: str) -> tuple[str, str] | None:
-    """Convert a full path back to a type-relative path.
-
-    Args:
-        full_path: The full path to the file
-
-    Returns:
-        tuple[str, str] | None: A tuple of (model_type, relative_path) if found, None otherwise
-    """
-    global folder_names_and_paths
-    full_path = os.path.normpath(full_path)
-
-    for model_type, (paths, _) in folder_names_and_paths.items():
-        for base_path in paths:
-            base_path = os.path.normpath(base_path)
-            if full_path.startswith(base_path):
-                relative_path = os.path.relpath(full_path, base_path)
-                return model_type, relative_path
-
-    return None
-
 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

main.py

@@ -112,6 +112,7 @@ import gc
 
 
 if os.name == "nt":
+    os.environ['MIMALLOC_PURGE_DELAY'] = '0'
     logging.getLogger("xformers").addFilter(lambda record: 'A matching Triton is not available' not in record.getMessage())
 
 if __name__ == "__main__":
@@ -164,6 +165,7 @@ def cuda_malloc_warning():
         if 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_instance):
     current_time: float = 0.0
     cache_type = execution.CacheType.CLASSIC
@@ -278,6 +280,7 @@ def cleanup_temp():
     if os.path.exists(temp_dir):
         shutil.rmtree(temp_dir, ignore_errors=True)
 
+
 def setup_database():
     try:
         from app.database.db import init_db, dependencies_available
@@ -286,6 +289,7 @@ def setup_database():
     except Exception as e:
         logging.error(f"Failed to initialize database. Please ensure you have installed the latest requirements. If the error persists, please report this as in future the database will be required: {e}")
 
+
 def start_comfyui(asyncio_loop=None):
     """
     Starts the ComfyUI server using the provided asyncio event loop or creates a new one.