FlashVSR

Project page arXiv paper Official code

FlashVSR is a one-diffusion-step streaming diffusion framework for real-time video super-resolution (VSR). It combines a train-friendly three-stage distillation pipeline, locality-constrained sparse attention that bridges the train-test resolution gap, and a tiny conditional decoder for fast reconstruction.

FlashVSR teaser figure.

Teaser image source: FlashVSR official repository.

Quick Start

V2V

uv sync --package flashdreams-flashvsr --inexact
uv run --no-sync flashdreams-run-v2 \
  v2v-flashvsr-v1.1-sparse-ratio-2.0 \
  --output-path artifacts/flashvsr.mp4 --timeout unbound -- --video-path input.mp4

Demo Presets Demo Arguments

Demo Presets

V2V

Preset	Description
`v2v-flashvsr-v1.1-sparse-ratio-2.0`	Stable sparse-attention 2x video super-resolution.
`v2v-flashvsr-v1.1-sparse-ratio-1.5`	Faster sparse-attention 2x video super-resolution.
`v2v-flashvsr-v1.1-full-attn`	Dense full attention with multi-GPU context-parallel support.

Developer Details

Integration Source Pipeline configurations

Minimum VRAM: about 24 GB.
Python: 3.10 or newer.
PyTorch: 2.9 or newer.

Post-processing presets

Cam2V and Interactive Drive commands select these presets with --postprocess-preset PRESET in their application arguments, after --:

Preset	Description
`flashvsr-v1.1-sparse-2.0`	N/A
`flashvsr-v1.1-sparse-1.5`	N/A
`flashvsr-v1.1-full-attn`	N/A

Performance (Outdated)

This historical benchmark compares per-chunk 2x upsampling time for FlashDreams FlashVSR with the official FlashVSR implementation under matched settings.

This chart records per-chunk 2x upsampling time in milliseconds on a single GB200 GPU with a chunk size of 8 frames. The current v2 application presets use 16-frame steady-state chunks, so these historical measurements are not timings of the current defaults. For the official FlashVSR implementation, see this instruction.

Samples

FlashVSR 2x output (1280x768) from flashvsr-v1.1-sparse-ratio-2.0; low-resolution input (672x384) inset at bottom-left. Input from the FlashVSR examples.

Citation

If you use FlashVSR, please cite the original work:

@inproceedings{zhuang2026flashvsr,
  title={FlashVSR: Towards Real-time Diffusion-Based Streaming Video Super Resolution},
  author={Zhuang, Junhao and Guo, Shi and Cai, Xin and Li, Xiaohui and Liu, Yihao and Yuan, Chun and Xue, Tianfan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={43482--43493},
  year={2026}
}