FlashVSR

FlashVSR

A one-step diffusion model that upscales low-resolution video as a stream, at about 17 frames per second for 768 x 1408 video on one A100 GPU. This is the first release, version 1.

Made by JunhaoZhuang on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: FlashVSR is a video super-resolution model. It takes low-resolution video and rebuilds it at higher resolution using a diffusion model that needs only one step, so it can work on a stream in close to real time.

The authors report about 17 frames per second for 768 x 1408 video on a single A100 GPU. They also built VSR-120K, a training dataset with 120k videos and 180k images, and say it scales to ultra-high resolutions.

Who it suits: Researchers and engineers who restore or upscale video and want a diffusion-based method that is fast enough for streaming. Creators who already use video tools with GPUs may also try it.

Industry | Who uses it | How they use it | Example
Research | Video restoration researchers | Compare one-step diffusion with older methods | Benchmarking against earlier diffusion upscalers
Media | Video editors | Upscale old or low-resolution footage | Sharpening archive clips
Streaming | Platform engineers | Enhance video in near real time | Upscaling a live feed on a GPU server
Creative | Hobbyists | Improve generated or downloaded clips | Raising a short clip to higher resolution

You can skip it if you have no GPU, or if you want a one-click app, since this is a research project with code in a separate repository.

Where to use it: The model is meant for GPU machines running the project's inference code. Community ports exist for tools like ComfyUI, and the card warns that some of them lose quality.

• GPU workstations and servers, with an A100 used for the reported speed
• The official inference code in the project's GitHub repository
• Third-party ports, with the quality caveat in the card

Use case | Who | What they get back
Video upscaling | Editors | Higher-resolution frames from low-resolution input
Streaming enhancement | Engineers | Near real-time super-resolution
Research baselines | Academics | A fast one-step diffusion method to compare

How to set it up: You need Python and a capable GPU. The card says the inference code and weights are released, and the visible part of its Getting Started section lists no commands, so follow the project's GitHub steps.

• Python with pip
• An NVIDIA GPU, with an A100 used in the reported speed
• The project's inference code from GitHub
• The Block-Sparse Attention library that the official pipeline relies on

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Follow the GitHub repository's steps to run the inference code.

pip install -U huggingface_hub
hf download JunhaoZhuang/FlashVSR

How to use it: Run the project's inference code on a low-resolution video and collect the upscaled output. The card gives no command in the visible part, so take the exact steps from the repository.

Example: You feed a short low-resolution clip into the official pipeline on a GPU. It returns an upscaled version of the clip, and the card says the official code keeps fine structures and avoids texture aliasing.

For newcomers and experts: If you are new, think of it as an upscaler for video that runs fast because it needs only one pass. Start with a short clip and compare the result with your original.

If you know the field, FlashVSR combines a three-stage distillation pipeline for streaming, locality-constrained sparse attention to close the train-test resolution gap, and a tiny conditional decoder. The authors report up to about 12 times speedup over earlier one-step diffusion VSR models.

Low-resolution video streams in
Locality-constrained sparse attention
One-step diffusion denoising
Tiny conditional decoder
High-resolution frames out

Good to know: Some third-party versions skip the locality-constrained sparse attention and fall back to dense attention, which the authors say can visibly lower quality at higher resolutions. A newer version 1.1 is a separate release, and this page is version 1.

• License: Apache 2.0
• Cost to run: none named, but the reported speed used an A100 GPU
• Data it reaches: the video you process, and the model files you download
• Not for: third-party ports without the sparse attention module if you need official quality

On Hugging Face, 13 November 2026

• Task: video to video
• Downloads in the last 30 days: 10,832
• Likes: 207
• License: apache-2.0
• Fine-tunes: 1
• Tags: flashvsr, video-to-video

Links

• Model: https://huggingface.co/JunhaoZhuang/FlashVSR
• Files: https://huggingface.co/JunhaoZhuang/FlashVSR/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download JunhaoZhuang/FlashVSR

Then follow the usage in its model card. Nothing in this kit runs by itself.