One PyTorch toolbox to upscale, denoise, and deblur them all

What it does BasicSR is an open-source PyTorch toolbox for image and video restoration. It collects implementations of many influential models spanning super-resolution, denoising, deblurring, and JPEG artifact removal. The project also folds in related utilities like StyleGAN2 training and face-restoration helpers such as DFDNet.

The interesting bit Instead of letting each paper live in its own brittle fork, BasicSR centralizes them under one configuration and dataset pipeline. That is why downstream projects like Real-ESRGAN and GFPGAN build on top of it, making it a de facto standard library for academic restoration research.

Key highlights

• Bundles ESRGAN, SwinIR, BasicVSR, EDVR, ECBSR, and many others in one codebase
• Supports both image and video restoration tasks (super-resolution, denoising, deblurring, JPEG removal)
• Provides plot scripts, Colab notebooks, and a separate BasicSR-Examples repo for package usage
• Apache 2.0 licensed and actively maintained by the XPixel group
• Underpins popular production projects like Real-ESRGAN and GFPGAN

Caveats

• Documentation for training and testing many core models (EDSR, ESRGAN, SRResNet, RCAN, EDVR, BasicVSR) is still marked as TODO in the HOWTOs table
• Several models only provide inference pipelines, not training code (DFDNet, DeblurGANv2, RIDNet, CBDNet)
• The README explicitly warns that quick-start pipelines do not cover all cases

Verdict Researchers and practitioners working on image or video restoration who want a unified PyTorch baseline should start here. Casual users looking for a polished one-click app should probably reach for Real-ESRGAN or GFPGAN instead.