One PyTorch repo collecting a decade of image restoration papers

What it does KAIR is a PyTorch toolbox that bundles implementations and training scripts for a wide range of image restoration architectures. It covers single-image super-resolution (SRMD, ESRGAN, BSRGAN, SwinIR), blind denoising (DnCNN, FFDNet, SCUNet), video restoration (VRT, RVRT), and even face enhancement (GPEN). Instead of scattered one-off repos, you get JSON-configured training loops, model summary utilities, and support for both single and multi-GPU setups in one place.

The interesting bit The repo acts less like a novel framework and more like a curated garage for the ETH Zurich lab’s own published work and related contemporaries. The unusual value is the sheer scope: classical CNNs, GAN-based perceptual methods, Swin Transformers, and video transformers all share a common training backbone. That makes it a practical archaeological layer cake of how image restoration research evolved from 2017 to 2023.

Key highlights

• Bundles code for DnCNN, FFDNet, ESRGAN, BSRGAN, SwinIR, USRNet, VRT, RVRT, and others
• Supports both PSNR-oriented and adversarial (GAN) training with DataParallel or DistributedDataParallel
• Includes utility scripts for FLOPs, parameter counts, receptive field calculation, and pretrained model downloads
• Covers tasks beyond single-image: video super-resolution/deblurring/denoising and face image enhancement
• JSON-based configuration files control dataset paths, GPU selection, and training modes

Caveats

• The README reads like a lab blogroll—useful for tracking updates, but architectural details are often deferred to separate paper repos
• Some newer methods (e.g., DiffPIR) are only linked externally and not actually integrated into the toolbox
• A 2020 correction note still warns that multi-GPU GAN training requires manually editing model_gan.py to comment out a line

Verdict Researchers actively reproducing or extending papers from this group will find it indispensable. Casual users looking for a simple pip-installable CLI should look elsewhere—this is a research workspace, not a polished end-user tool.