mv-lab/swin2sr

Based on the README, here's a technical summary: Built on SwinV2 Transformer architecture with improved training stability and faster convergence (~2x speedup over SwinIR), Swin2SR addresses resolution gaps between pre-training and fine-tuning while reducing data hunger—key challenges in vision transformer training. The model handles multiple restoration tasks including JPEG artifact removal, classical/lightweight super-resolution, and compressed image upscaling (4x scale by default). Integrated into Stable Diffusion WebUI and deployable via PyTorch with pretrained weights on DIV2K/Flickr2K datasets; inference code and comprehensive visual results are provided in the repository.

676 stars. No commits in the last 6 months.