FlashVSR+

FlashVSR+

A one-step diffusion model that sharpens low-quality video into higher-resolution video, with a fast Turbo variant and a higher-quality Base variant. For video editors and vision researchers with a capable GPU.

Made by ruchangyao 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 a low-quality clip and returns a sharper, higher-resolution one through a one-step diffusion process, so one pass replaces many. It builds on Wan2.1-T2V-1.3B. The authors say it can handle 4K video on one RTX 4090 with 24GB of memory without cutting frames into tiles.

Who it suits: People who restore or upscale video and have a capable GPU, plus researchers who study efficient diffusion for video.

Industry | Who uses it | How they use it | Example
Video production | Editors restoring footage | Upscale low-resolution clips | Turn a soft clip into a sharper high-resolution version
Computer vision research | Researchers on diffusion for video | Study one-step inference and attention design | Compare the Turbo and Base outputs on the same clip
Media archives | People digitizing older footage | Raise the resolution of stored video | Run an old low-quality recording through the model

People without a capable GPU can skip it, since the 4K claim assumes a 24GB RTX 4090.

Where to use it:

• A computer with a strong CUDA GPU, such as the 24GB RTX 4090 the authors name
• Windows, macOS or Linux with Python, to download the files
• The project's own code repository, for running inference

Use case | Who | What they get back
Restore low-quality video | Video editors | A sharper clip at a higher resolution
Fast upscaling with Turbo | People with many clips | Quicker inference for high-resolution work
Quality-first restoration with Base | Anyone who values reconstruction quality | Higher-quality restoration from the same one-step setup

How to set it up: You need Python with pip and a GPU of the class the authors quote. The checkpoints sit in folders named turbo and base, with a shared decoder beside them.

• Python and pip
• Enough disk space for the checkpoints
• A CUDA GPU, with 24GB of memory for the 4K case

1. Install the Hugging Face hub package.
2. Download the repository.
3. Choose the turbo or base folder for your job.

pip install -U huggingface_hub
hf download ruchangyao/FlashVSR_plus

How to use it: The model card lists the released files but shows no run command. Pair the checkpoint folder you chose with the project's inference code, then feed it a low-quality video.

Tool | What it does
FlashVSR+ Turbo | Faster inference, suggested for efficient high-resolution video
FlashVSR+ Base | Higher-quality restoration when reconstruction quality matters most
TCDecoder.ckpt | The decoder both variants share in the default pipeline
Wan2.1_VAE.pth | Optional VAE decoder weights

Example: You have a low-resolution clip and want a sharper copy quickly. You pick the Turbo checkpoint and run the clip through the project's inference code. A higher-resolution version comes back, made in a single step. The card shows no exact command for this.

For newcomers and experts: Video super-resolution means making a small, blurry video larger and clearer. A good first step is to download the files and look at the turbo and base folders to see what each variant ships.

For experts, the speed comes from hybrid attention, which uses locality in space and time and keeps a few sparse global links. The authors quote up to 2x faster DiT and 45% lower peak memory. Their RY-DMD distillation step tunes the model toward better perceptual quality using reward scores. The cost is a choice between Turbo for speed and Base for quality.

Low-quality video in
Hybrid attention DiT, one step
Shared TCDecoder
Higher-resolution video out

Good to know:

• Hardware: the 4K figure is stated for one 24GB RTX 4090
• Files: separate turbo and base checkpoints plus a shared decoder
• Claims: speed and memory gains come from the authors' own write-up
• License: Apache 2.0
• Cost to run: no API key or paid service named, but it needs a capable GPU
• Data it reaches: downloads files from Hugging Face, no account or other service is named
• Not for: the card names no limits, and it is built for video super-resolution

On Hugging Face, 13 November 2026

• Task: video to video
• Downloads in the last 30 days: 1,269
• Likes: 36
• License: apache-2.0
• Based on: Wan-AI/Wan2.1-T2V-1.3B
• Tags: video-super-resolution, video-restoration, flashvsr, video-to-video

Links

• Model: https://huggingface.co/ruchangyao/FlashVSR_plus
• Files: https://huggingface.co/ruchangyao/FlashVSR_plus/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ruchangyao/FlashVSR_plus

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