# OpenImagingLab/FlashVSR

[CVPR 2026] Towards Real-Time Diffusion-Based Streaming Video Super-Resolution — An efficient one-step diffusion framework for streaming VSR with locality-constrained sparse attention and a tiny conditional decoder.

Repository: https://github.com/OpenImagingLab/FlashVSR
Canonical: https://ross.abutalabs.com/products/flashvsr
Homepage: https://zhuang2002.github.io/FlashVSR/
Language: Python
License: Apache-2.0
License Family: permissive
Topics: diffusion-models, video-super-resolution, video-restoration
Last push: 2026-08-23T17:45:39+00:00

## Health v2 (maintenance only)
Score: 61/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 35, longevity 23
- inputs: {"age_days": 323, "days_push": 10, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1799, forks 145 (observed 2026-08-28T04:05:37.865742+00:00)

## What it is
FlashVSR is a one-step diffusion-based streaming video super-resolution framework that runs at ~17 FPS for 768x1408 video on a single A100 GPU. It combines a three-stage distillation pipeline, locality-constrained sparse attention, and a tiny conditional decoder, with pretrained model weights and a VSR-120K training dataset.

## Use cases
- upscale low-resolution videos in real time
- restore and enhance blurry or degraded video streams
- run diffusion-based video super-resolution efficiently on a single GPU
- super-resolve ultra-high-resolution videos
- integrate video super-resolution into ComfyUI workflows
- train video super-resolution models on a large-scale dataset

## When to choose
- you need real-time or streaming video super-resolution rather than offline batch processing
- you want diffusion-quality restoration with far lower latency than multi-step diffusion VSR models
- you need to scale super-resolution to ultra-high resolutions without artifacts
- you want pretrained weights and a ComfyUI integration for quick use

## When to avoid
- you only need simple classical upscaling (e.g., bicubic or ESRGAN) with minimal GPU cost
- you lack a CUDA-capable GPU, since the framework targets A100-class hardware for real-time speeds
- you need image-only super-resolution without video/temporal handling
- you require a permissively simple setup without diffusion model dependencies

## Facets
- artifact type: library
- maturity: active
- function: video-processing, image-processing, machine-learning, deep-learning, llm-inference
- domain: computer-vision, image-processing, deep-learning, artificial-intelligence
- platform: python, cross-platform
- tags: diffusion-models, video-super-resolution, video-restoration, one-step-diffusion, streaming-inference, sparse-attention, cvpr-2026, comfyui, video, gpu, linux

## Member repositories
- OpenImagingLab/FlashVSR (main) score 61

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:37.865742+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:22:16.307617+00:00, confidence not recorded.
  - readme: https://github.com/OpenImagingLab/FlashVSR (fetched 2026-08-28T04:05:37.865742+00:00, sha 4deda4f1d145)
  - homepage: https://zhuang2002.github.io/FlashVSR/ (fetched 2026-08-29T11:01:07.565532+00:00, sha 99c49cddf9ee)
- Data as of 2026-08-30T08:39:29.467469+00:00.