RelayVSR

RelayVSR is a research framework for real-world video super-resolution from the University of Science and Technology of China (arXiv 2026). It uses a "Sparse Generative Relay": a large generative model produces reference latents for sparse keyframes, and a lightweight network, a Dual-Memory Video Transformer, uses those references plus the low-resolution video to super-resolve every frame.

The lightweight network reuses keyframe information across frames and updates recent video context. It supports first-keyframe conditioning and dual-endpoint conditioning with bounded lookahead. Training uses Video-Aware Reference Optimization (VARO), a reinforcement-learning step that updates the large model with a system-level reward on the final video and a reference-level reward on decoded keyframes.

The authors report that at 1080p on a single NVIDIA A100 80GB, the dual-endpoint variant with a 15-frame keyframe interval reaches 29.29 FPS, 13.82 GB peak GPU memory and 0.327 s first-frame latency. They give 7.80 FPS, 24.447 GB and 2.83 s for FlashVSR-Tiny as the comparison. A YouTube creator described it as a 4x upscaler built on Alibaba's Wan, about 10.2 GB, though this was not confirmed on the project page. The page says code is available on GitHub, but no repository, weights or license link was provided in the material reviewed.