RelayVSR is a new open model for real-world video super-resolution: give it a blurry, low-resolution clip and it returns the same video four times larger, with sharper detail and stable motion. It pairs a large Wan video model, which renders a reference frame every so often, with a small fast model that fills in the frames between. The code and weights are Apache-2.0, but running it needs a large GPU and a CUDA setup. On LiveWeights you just upload the video.
RelayVSRUSTC (Xijun Wang et al.) · Apache-2.0Real-world video super-resolution: upload a low-resolution clip and get it back 4× sharper, with the original audio.Try it online →How RelayVSR works
- Large and small model in relay: the Wan model produces high-quality reference frames every 15 frames; the compact Flash model upscales the frames in between using those references.
- Streaming: frames are processed in order, so long videos don't need all frames in memory at once.
- Real-world footage: trained for compressed, noisy, blurry video like old clips and phone recordings, not just clean downscaled images.
License: can I use it commercially?
RelayVSR is released under Apache-2.0, so commercial use is allowed. Your upscaled videos are yours.
How to try it online
- Open the RelayVSR page and sign in, then pick a small credit pack (no subscription).
- Upload a short video (MP4, MOV or WebM, up to 8 MB). Up to 10 seconds are processed.
- Press Upscale 4×. Inputs larger than 540 px on the long side are scaled down first, so the result is at most 2160 px. The original audio is kept.
- Switch between Upscaled and Original to compare, then download the MP4.
Tips for better results
- The smaller the input, the bigger the improvement: 240p to 480p clips gain the most.
- Upscaling can't recover details that aren't there at all, like unreadable distant text, but it cleans up edges, faces and textures.
- Try it on clips from AI video models too: many render at low resolution.
Run it locally
The inference code is on GitHub (kopperx/RelayVSR) and the two checkpoints on Hugging Face (kopper/RelayVSR, about 10 GB). It needs Linux, Python 3.11+, PyTorch 2.10 and an NVIDIA GPU. Comfy-Org also publishes repackaged files for ComfyUI, and LightX2V a 2× version.