KAIR

Why we included this project

KAIR is a practical PyTorch toolbox for image restoration that collects training and evaluation code for a broad set of published models. Denoising networks like DnCNN and FFDNet, super-resolution approaches such as SRMD, BSRGAN, and SwinIR, and video restoration methods like VRT and RVRT all live under one coherent codebase. It is not a single drop-in application but a reference implementation library, so you can train a network from scratch on your own data or load pretrained weights and run inference on noisy, blurry, or low-resolution images. The bundled scripts also report FLOPs, parameter counts, and memory usage, which helps when comparing architectures for deployment constraints. Because it covers so many tasks, it is a good starting point for adapting a published restoration method instead of reimplementing it from scratch.