BasicVSR++

BasicVSR++ (CVPR 2022, arXiv 2104.13371) is the video super-resolution model from Kelvin C.K. Chan's team at NTU that became the field's default baseline almost overnight. It extends BasicVSR with two ideas: a second-order grid propagation scheme, which lets features circulate through more paths between past and future frames, and flow-guided deformable alignment, which fuses optical-flow guidance with deformable offsets so alignment stays stable where plain deformable convolutions drift. The release also includes training strategies that matter in practice and 2x, deblur, and denoise variants.

The code lives at github.com/ckkelvinchan/BasicVSR_PlusPlus with pretrained checkpoints, a Colab demo, and integration into OpenMMLab's MMEditing. It is a research artifact, not a product: there is no API, no hosted service, and no commercial licensing framework beyond the repo's research terms. Practically everything in open-source video SR since 2022 benchmarks against it, which is the strongest signal of its standing.

Author-reported (CVPR 2022): state-of-the-art on REDS4 and Vimeo-90K at publication; widely adopted as the comparison baseline by subsequent VSR work