FlashVSR: Real-Time Video Super-Resolution for Wan2.1
FlashVSR by JunhaoZhuang is a real-time video super-resolution model for Wan2.1. Features LQ projection model and TC decoder as model_patches for streaming VSR at ~17 FPS.
FlashVSR
Wan2.1Video Super-ResolutionReal-TimeUpscalingReal-time video super-resolution model for Wan2.1. Uses LQ projection model and TC decoder as model_patches for streaming diffusion-based VSR at approximately 17 FPS on A100.
| Developer | JunhaoZhuang (OpenImagingLab) |
| Architecture | Wan2.1-based streaming VSR |
| Model Type | Video super-resolution (model_patches) |
| Patches | LQ projection model, TC decoder |
| Performance | Approximately 17 FPS on A100 (768x1408) |
| Conference | CVPR 2026 |
Overview
FlashVSR (CVPR 2026) is a real-time diffusion-based streaming video super-resolution model built on Wan2.1. It achieves approximately 17 FPS for 768x1408 videos on a single A100 GPU through three key innovations: a train-friendly three-stage distillation pipeline for streaming super-resolution, locality-constrained sparse attention, and a tiny conditional decoder.
The model is implemented as model_patches for Wan2.1, featuring an LQ (low-quality) projection model and a TC (tiny conditional) decoder that can be loaded alongside the base Wan2.1 model.
ComfyUI Integration
FlashVSR patches are compatible with ComfyUI via Kijai's WanVideoWrapper. The LQ projection model and TC decoder are loaded as model_patches on top of the base Wan2.1 model. Model files are available on the Kijai/WanVideo_comfy repository.
Guides and workflows related to this model series.