ComfyUI-FlashVSR_Ultra_Fast

Running FlashVSR on lower VRAM without any artifacts.
[📃中文版本]

Changelog

2025-10-24

• Added long video pipeline that significantly reduces VRAM usage when upscaling long videos.

2025-10-21

• Initial this project, introducing features such as tile_dit to significantly reducing VRAM usage.

2025-10-22

• Replaced Block-Sparse-Attention with Sparse_Sage, removing the need to compile any custom kernels.
• Added support for running on RTX 50 series GPUs.

Preview

Usage

• mode:
  tiny -> faster (default); full -> higher quality
• scale:
  4 is always better, unless you are low on VRAM then use 2
• color_fix:
  Use wavelet transform to correct the color of output video.
• tiled_vae:
  Set to True for lower VRAM consumption during decoding at the cost of speed.
• tiled_dit:
  Significantly reduces VRAM usage at the cost of speed.
• tile_size, tile_overlap:
  How to split the input video.
• unload_dit:
  Unload DiT before decoding to reduce VRAM peak at the cost of speed.

Installation

nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/lihaoyun6/ComfyUI-FlashVSR_Ultra_Fast.git
python -m pip install -r ComfyUI-FlashVSR_Ultra_Fast/requirements.txt

models:

• Download the entire FlashVSR folder with all the files inside it from here and put it in the ComfyUI/models

├── ComfyUI/models/FlashVSR
|     ├── LQ_proj_in.ckpt
|     ├── TCDecoder.ckpt
|     ├── diffusion_pytorch_model_streaming_dmd.safetensors
|     ├── Wan2.1_VAE.pth

Acknowledgments

• FlashVSR @OpenImagingLab
• Sparse_SageAttention @jt-zhang
• ComfyUI @comfyanonymous