cszn/KAIR

Supports both image and video restoration tasks across super-resolution, deblurring, and denoising with modular architecture enabling plug-and-play restoration. Implements transformer-based models (SwinIR, VRT, RVRT) alongside convolutional architectures, with distributed GPU training and comprehensive degradation simulation for blind real-world restoration. Includes pretrained models, perceptual loss variants, PatchGAN discriminators, and integrated benchmarking utilities for FLOPs and memory profiling.

3,444 stars. No commits in the last 6 months.