CVPR19’s video-restoration winner now lives in a broader toolbox

What it does BasicSR is a PyTorch toolbox for image and video restoration research. It packages training and inference code for super-resolution, deblurring, denoising, JPEG artifact removal, and even blind face restoration. The codebase also implements several well-known models, including the NTIRE19-winning EDVR, ESRGAN, and StyleGAN2.

The interesting bit EDVR originally introduced enhanced deformable convolutions for video restoration, but the authors have since folded it into BasicSR as a sub-component. This repository is now just a mirror of that larger framework, so what was once a standalone champion has become one module in a general-purpose restoration factory.

Key highlights

• Took first place in the NTIRE19 video restoration and enhancement challenges at CVPR19 workshops.
• Bundles multiple restoration tasks—super-resolution, denoising, deblurring, and face restoration—under one training pipeline.
• Publishes pre-trained models and training curves via Weights & Biases.
• Requires custom CUDA extensions for EDVR and StyleGAN2; compilation is mandatory for those features.
• Released under the Apache 2.0 license.

Caveats

• This repo is a mirror of BasicSR and breaks compatibility with previous EDVR versions; legacy code is stranded on the old_version branch.
• The developers only test on Ubuntu and explicitly caution that Windows is unsupported.
• Several HOWTO documentation entries remain marked as TODO.

Verdict A solid starting point for researchers who want a unified PyTorch framework to benchmark or extend image and video restoration models. If you just need the original, minimal EDVR inference code, the old_version branch is a better bet than this mirror.