ComfyUI-WanVideoWrapper

This project only supports NVIDIA RTX30/40 and A100/H100 GPU!

This is a modified version of ComfyUI-WanVideoWrapper.
I added draft-attention support and Euler/d scheduler to ComfyUI-WanVideoWrapper to further speed up video inference.
And this update significantly improves the inference speed of NVIDIA Ampere / Ada / Hopper GPU.
The average inference speed is increased by more than 25~40% compared to sage-attn on any Ampere GPU.
[中文版本] [original readme]

Demo

Input Image	Sage Attention	Draft Attention (75% sparsity)	Draft Attention (90% sparsity)
Sampling time:	70.38s	48.31s	40.91s

Prompt: "The girl raised her hand to block the sun, with a charming smile on her face, and the camera gradually zoomed out"

Usage

After installing the backend library and ComfyUI plugin, you can use it just like the original WanVideoWrapper. But unlike other attention mode, draft-attention does not currently support arbitrary resolution video generation.
So you need to use the Find Draft-Attention Bucket node to convert the video size you set to a size that draft-attention can handle.

Installation

! Please uninstall the original Comfyui-WanVideoWrapper before installing my version!

Pre-Build Installation (✅Recommend):

1. Clone this repo and install the dependencies:

   cd [ComfyUI_Path]\custom_nodes
   git clone https://github.com/lihaoyun6/ComfyUI-WanVideoWrapper
   pip install -r ComfyUI-WanVideoWrapper\requirements.txt
   

2. Install flash-attention and Block-Sparse-Attention.

   There are only wheels for Windows_x64 here!
   It's recommended to update GPU driver and install torch2.7.0+cu128, then install wheels according to your python version:

   # Python 3.10
   pip install https://github.com/lihaoyun6/Block-Sparse-Attention/releases/download/v0.0.1/flash_attn-2.8.1+cu128torch2.7cxx11abiFALSE-cp310-cp310-win_amd64.whl
   pip install https://github.com/lihaoyun6/Block-Sparse-Attention/releases/download/v0.0.1/block_sparse_attn-0.0.1+cu128torch2.7cxx11abiFALSE-cp310-cp310-win_amd64.whl
   

   # Python 3.11
   pip install https://github.com/lihaoyun6/Block-Sparse-Attention/releases/download/v0.0.1/flash_attn-2.8.1+cu128torch2.7cxx11abiFALSE-cp311-cp311-win_amd64.whl
   pip install https://github.com/lihaoyun6/Block-Sparse-Attention/releases/download/v0.0.1/block_sparse_attn-0.0.1+cu128torch2.7cxx11abiFALSE-cp311-cp311-win_amd64.whl
   

   # Python 3.12
   pip install https://github.com/lihaoyun6/Block-Sparse-Attention/releases/download/v0.0.1/flash_attn-2.8.1+cu128torch2.7cxx11abiFALSE-cp312-cp312-win_amd64.whl
   pip install https://github.com/lihaoyun6/Block-Sparse-Attention/releases/download/v0.0.1/block_sparse_attn-0.0.1+cu128torch2.7cxx11abiFALSE-cp312-cp312-win_amd64.whl
   

   # Python 3.13
   pip install https://github.com/lihaoyun6/Block-Sparse-Attention/releases/download/v0.0.1/flash_attn-2.8.1+cu128torch2.7cxx11abiFALSE-cp313-cp313-win_amd64.whl
   pip install https://github.com/lihaoyun6/Block-Sparse-Attention/releases/download/v0.0.1/block_sparse_attn-0.0.1+cu128torch2.7cxx11abiFALSE-cp313-cp313-win_amd64.whl
   

Manual Installation:

1. Clone this repo and install the dependencies:

   cd [ComfyUI_Path]\custom_nodes
   git clone https://github.com/lihaoyun6/ComfyUI-WanVideoWrapper
   pip install -r ComfyUI-WanVideoWrapper\requirements.txt
   

2. Build and install Block-Sparse-Attention:

   Please make sure you have installed the CUDA>=11.6, C++ compiler and PyTorch

   pip install ninja wheel packaging
   #Each JOB will use about 8GB of RAM, please set it appropriately.
   #For example, if the you have 64GB RAM, you can set MAX_JOBS=8
   export MAX_JOBS=4
   #Download source code and install.
   #This process may take 1~3 hours depending on your CPU and RAM. Please be patient.
   git clone https://github.com/lihaoyun6/Block-Sparse-Attention
   cd Block-Sparse-Attention
   pip install --no-build-isolation ./
   

3. Install flash-attention:

   pip install flash-attention --no-build-isolation
   

Thanks

ComfyUI-WanVideoWrapper - kijai
draft-attention - shawnricecake
Block-Sparse-Attention - mit-han-lab
Euler/d - eddyhhlure1Eddy