SparkVSR

Why we included this project

Most video upscalers run as a black box: you feed in low-res footage and take whatever comes out. SparkVSR gives you a handle on that process. You pick a small set of keyframes, upscale them with any image super-resolution model you like, and the network spreads those results across the rest of the clip while staying true to the original motion. Because you choose which frames to refine, manually, from codec I-frames, or at random, you can steer the output where a one-shot model can't, and it tolerates imperfect keyframes rather than falling apart. The same pipeline also handles old-film restoration and video style transfer, and a ComfyUI node makes it easy to slot into an existing editing setup, with training and inference code, pretrained weights, and benchmark results all in the repo.