Xijun Wang, Xin Li, Zirui Lang, Suhang Yao, Haoran Li, Zhibo Chen (University of Science and Technology of China) · Streaming video super-resolution

RelayVSR

RelayVSR is a streaming real-world video super-resolution framework from the University of Science and Technology of China. It uses a Sparse Generative Relay mechanism: a large generative model produces reference latents for sparse keyframes, and a lightweight Dual-Memory Video Transformer uses those references and the low-resolution video to super-resolve every frame. The lightweight network supports first-keyframe conditioning and dual-endpoint conditioning with bounded lookahead. Video-Aware Reference Optimization (VARO) uses reinforcement learning to update the large model with a system-level reward on the output video and a reference-level reward on decoded keyframe quality. The authors say VARO improves on direct joint training and that dual-level rewards beat a system-level reward alone. At 1080p on a single A100 80GB, dual-endpoint RelayVSR with a 15-frame keyframe interval reaches 29.29 FPS, 13.82 GB peak GPU memory and 0.327 s first-frame latency, against 7.80 FPS, 24.447 GB and 2.83 s for FlashVSR-Tiny. Code and models are linked from the project page (arXiv 2026).