Repolode
Submit
OpenImagingLab

OpenImagingLab/FlashVSR

A one-step diffusion-based streaming video super-resolution framework aimed at real-time upscaling, built on Python with Wan-based models.

View on GitHubWebsite
Stars
1,934
Forks
163
Open issues
78
Last commit
4 days ago

What is it?

FlashVSR is a video super-resolution project that uses diffusion models to upscale video frames in a streaming fashion. It is designed for real-world footage where earlier diffusion-based VSR approaches are too slow and too memory-hungry to run in near real time, especially at ultra-high resolutions.

The framework combines a three-stage distillation pipeline that enables one-step streaming inference, locality-constrained sparse attention that reduces redundant computation, and a small conditional decoder for faster reconstruction. The authors report roughly 17 FPS for 768 × 1408 videos on a single A100 GPU. The repository includes inference scripts for v1 and v1.1 model weights hosted on Hugging Face, with full and tiny variants plus a long-video script. It is primarily optimized for 4× upscaling.

Setup requires Python 3.11, a build of the Block-Sparse Attention backend, and Git LFS to fetch weights. The project is Apache-2.0 licensed.

Key features

  • 01One-step diffusion inference designed for streaming video super-resolution.
  • 02Locality-constrained sparse attention to reduce computation at high resolutions.
  • 03Tiny conditional decoder for faster frame reconstruction.
  • 04Inference scripts for v1 and v1.1 weights, including full, tiny and long-video variants.
  • 05Primarily optimized for 4× upscaling.

Use cases

  • Upscaling low-resolution video footage to higher resolutions with diffusion-based quality.
  • Running video restoration experiments on A100-class GPUs.
  • Processing long videos frame by frame with a streaming pipeline.

Quick start

  1. 1

    Clone the repository

    git clone https://github.com/OpenImagingLab/FlashVSR
  2. 2

    Install Python dependencies (Python 3.11.13; also build Block-Sparse Attention separately from the README)

    pip install -r requirements.txt
  3. 3

    Download the recommended v1.1 model weights from Hugging Face with Git LFS

    git lfs clone https://huggingface.co/JunhaoZhuang/FlashVSR-v1.1
  4. 4

    Run inference from examples/WanVSR with the v1.1 tiny script

    python infer_flashvsr_v1.1_tiny.py

Things to know

  • Block-Sparse Attention is required and its build is memory-intensive.
  • Tested mainly on NVIDIA A100/A800 and H200; other GPUs such as RTX 40/50 series are unverified.
  • Primarily designed and optimized for 4× super-resolution only.

Topics

#video-super-resolution#diffusion-models#video-restoration#real-time#computer-vision#pytorch

FAQ

What is FlashVSR?
A one-step diffusion-based streaming video super-resolution framework aimed at real-time upscaling, built on Python with Wan-based models. FlashVSR is a video super-resolution project that uses diffusion models to upscale video frames in a streaming fashion. It is designed for real-world footage where earlier diffusion-based VSR approaches are too slow and too memory-hungry to run in near real time, especially at ultra-high resolutions.
Is FlashVSR free to use? What license does it have?
FlashVSR is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is FlashVSR still maintained?
Actively maintained. The last commit to FlashVSR was on Oct 16, 2026.
What can I use FlashVSR for?
Upscaling low-resolution video footage to higher resolutions with diffusion-based quality. Running video restoration experiments on A100-class GPUs. Processing long videos frame by frame with a streaming pipeline.
How do I install FlashVSR?
Clone the repository: git clone https://github.com/OpenImagingLab/FlashVSR Install Python dependencies (Python 3.11.13; also build Block-Sparse Attention separately from the README): pip install -r requirements.txt Download the recommended v1.1 model weights from Hugging Face with Git LFS: git lfs clone https://huggingface.co/JunhaoZhuang/FlashVSR-v1.1 Run inference from examples/WanVSR with the v1.1 tiny script: python infer_flashvsr_v1.1_tiny.py
Where can I get FlashVSR?
The source code and installation instructions are on GitHub: https://github.com/OpenImagingLab/FlashVSR

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects