RelayVSR: Generative 4x Video Upscaling at 29 FPS

Researchers from the University of Science and Technology of China introduced RelayVSR (arXiv:2609.37850) — a streaming 4x video upscaling framework where a large Wan-based model processes only rare keyframes, while a lightweight Dual-Memory Video Transformer handles the rest, achieving 29.29 FPS at 1080p on a single A100 with 13.82 GB of memory.

Compiled by Sergey KostenchukPublished 2026-11-06Updated 2026-11-06

2026-11-06 Research HuggingFace

🤖 RelayVSR: Generative 4x 1080p Video Upscaling at 29 FPS

The USTC team published RelayVSR (arXiv:2609.37850) — a streaming 4x upscaling framework. Sparse Generative Relay: a Wan-based model renders anchor latents only on rare keyframes, while a lightweight Dual-Memory Video Transformer fills in the rest; training uses RL (VARO) with rewards for the final output and keyframes.

🌍 Author measurements on a single A100 at 1080p: 29.29 FPS and 13.82 GB compared to 7.80 FPS and 24.4 GB for FlashVSR-Tiny. The large model is invoked only for a small fraction of frames, and RL maintains quality, so hardware requirements are ~3.75x lower in speed and ~43% lower in memory — making generative upscaling realistic for streaming and long videos.

👤 Code and weights are open under Apache-2.0, with a ComfyUI port available (2,607 downloads in one month): try 4x upscaling on real or generated video.

Source 1: https://arxiv.org/abs/2609.37850 Source 2: https://kopperx.github.io/RelayVSR/

Sources