RelayVSR

RelayVSR is a free open-source streaming 4× video super-resolution system: a large model guides sparse keyframes, a small network fills every frame.

What it is: large-model keyframes, small-model video

RelayVSR is from Xijun Wang, Xin Li, Zirui Lang, Suhang Yao, Haoran Li and Zhibo Chen at the University of Science and Technology of China. The paper, code and project page describe a streaming VSR system built around Sparse Generative Relay: a large generative model writes reference latents only on sparse keyframes, then a lightweight Dual-Memory Video Transformer uses those references plus the low-resolution video to reconstruct every frame.

• Fixed 4× only: output width and height are exactly four times the input. Frame rate is kept. Audio is not copied.
• Two conditioning modes: first-keyframe conditioning, or dual-endpoint conditioning with a bounded lookahead so the small model can use references at both ends of a window.
• VARO: Video-Aware Reference Optimization trains the large model with reinforcement learning. One reward scores the final video from the fixed small network; another scores decoded keyframe quality. The authors say this beats direct joint training, and that using both rewards beats a system-level reward alone. That comparison is from the paper, not a Seailife run.

What the paper reports on an A100

• At 1080p on one NVIDIA 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 model latency.
• The same table lists FlashVSR-Tiny at 7.80 FPS, 24.447 GB and 2.83 s. These numbers are the authors' measurements.

How to run the official repo

• Machine: Linux, Python 3.11–3.13 and an NVIDIA GPU. Setup is uv sync --locked --extra cu128, then activate the venv.
• Weights: download wan.safetensors and flash.safetensors from huggingface.co/kopper/RelayVSR into weights/. On November 12 those files were about 10.07 GB and 0.13 GB.
• Infer:python inference.py --output output_x4.mp4. Pass --input for another video, or an ordered image directory with --fps. The bundled demo clip is used when --input is omitted.
• Defaults called out in the README: Flash checkpoint S at step 60000 with EMA, a two-frame attention window, and --window-size 16 (reference endpoints every 15 frames). You can override checkpoints, dtype (bf16, fp16, fp32), --max-frames and --overwrite.

Related releases, and what this page covers

• The license is Apache-2.0. The README says the Wan implementation keeps its upstream copyright notice, and thanks Diffusers, Wan and FlashVSR.
• ComfyUI has built-in RelayVSR nodes, and LightX2V publishes a separate RelayVSR X2 checkpoint for native fixed 2× upscaling. Those are other packages. This page covers only the official 4× kopperx/RelayVSR release.
• The GitHub repo was created on September 29, 2026. On November 12 it had about 262 stars, 21 forks and 3 open issues and pull requests together. Hugging Face listed about 61 likes on the weight repo.
• Seailife did not install the env, download the weights or run inference. Speed and memory figures above come from the paper and project page. The logo is the project's own icon; the product image is the overview figure from the repository.