Video super-resolution often faces a choice between detailed results that take a long time to produce and fast processing that misses plausible detail. RelayVSR’s central idea is to spend generative-model compute on only a sparse set of keyframes, then let a smaller network carry that information across the rest of the video. This Comfy-Org repository packages the weights for use in ComfyUI rather than introducing a separate model architecture.
Key Capabilities
- The 2× and 4× variants target different upscaling needs, with matching VAE weights packaged for the ComfyUI workflow.
- Sparse references let the large generative model run intermittently while a compact video network handles intermediate frames, reducing repeated heavyweight inference.
- The underlying paper reports 29.29 FPS at 1080p with 13.82 GB peak GPU memory in dual-endpoint mode on one NVIDIA A100 80GB, using a 15-frame keyframe interval. These are paper-reported measurements, not a guarantee for every packaged variant or setup.
Who It’s For
A good fit for ComfyUI users who want generative detail recovery for video and can run the required models locally. Look elsewhere if you need audio-preserving output, a lightweight CPU workflow, or predictable frame-by-frame results without generative variation; the reference model’s inference pipeline does not copy audio, and results depend on GPU and configuration.
