SwinIR is the official implementation of “SwinIR: Image Restoration Using Swin Transformer”, a widely used model family for restoring and enhancing images. A single architecture handles super-resolution, denoising, JPEG artifact reduction and real-world image enhancement, and the repository provides pretrained models plus complete training and testing code. It has more than 5,600 stars on GitHub and remains a reference implementation for low-level vision.
Built on the Swin Transformer’s shifted-window attention, SwinIR set strong benchmarks across classical and real-world restoration tasks when it appeared, and its checkpoints have since been baked into countless photo-enhancement tools and pipelines. The repository covers classical and lightweight super-resolution, grayscale and colour denoising, compression artifact reduction and real-world super-resolution, with clear instructions for running the pretrained models or retraining on your own data. Its design keeps the strong restoration transformer backbone of earlier work while adding residual connections and training strategies that stabilise larger models, which is part of why it generalised well beyond its original benchmarks. Years after publication it is still frequently the first model developers reach for when an application needs dependable image enhancement. Its checkpoints are distributed through several model hubs and demo spaces, so trying it on a photo takes minutes, and its modular PyTorch code is straightforward to embed in larger vision systems or to fine-tune on domain-specific imagery such as medical or satellite photos.
Free. SwinIR is released under the Apache-2.0 license, including its pretrained models, for research and commercial use.
Researchers and developers working on image restoration, enhancement and super-resolution who need a proven transformer-based baseline or production component.