repository·main·Indexed 26 days ago
https://github.com/jingyunliang/swinirA PyTorch implementation of a high-performance image restoration framework based on the Swin Transformer. SwinIR supports multiple tasks including image super-resolution (classical, lightweight, and real-world), grayscale and color image denoising, and JPEG compression artifact reduction.
SwinIR is a PyTorch implementation of a Swin Transformer-based model for image restoration. It achieves state-of-the-art performance in the following tasks:
You can run SwinIR for several restoration tasks using the main_test_swinir.py script. The script supports classical super-resolution, lightweight super-resolution, real-world super-resolution, denoising (grayscale and color), and JPEG artifact reduction.
Key CLI Arguments:
--task: The restoration task (e.g., classical_sr, lightweight_sr, real_sr, gray_dn, color_dn, jpeg_car, color_jpeg_car).--scale: The upscaling factor (for SR tasks).--model_path: Path to the pretrained .pth model file.--folder_lq: Path to the folder containing low-quality input images.--folder_gt: Path to the folder containing ground-truth images (used for evaluation).--tile: Use this flag (e.g., --tile 400) if you encounter out-of-memory errors during real-world SR.--large_model: Use this flag for larger models in real-world SR.--noise: Specify the noise level (for denoising tasks).--jpeg: Specify the JPEG quality factor (for artifact reduction tasks).Pre-trained SwinIR models for various image restoration tasks can be downloaded from the official GitHub releases page.
https://github.com/JingyunLiang/SwinIR/releases/tag/v0.0Datasets required for training and testing vary by task. Below are the primary datasets used for each task:
If you do not want to set up a local environment, you can use the following interactive demos:
The training code for SwinIR is maintained within the KAIR repository. You can find the specific documentation and implementation details for SwinIR training there.
https://github.com/cszn/KAIR/blob/master/docs/README_SwinIR.mdThe main_test_swinir.py script uses the following arguments to control the testing process across different tasks:
| Argument | Description |
|---|---|
--task | Restoration task: classical_sr, lightweight_sr, real_sr, gray_dn, color_dn, jpeg_car, color_jpeg_car |
--scale | Upscaling factor for SR tasks |
--training_patch_size | Used to differentiate settings in classical SR |
--model_path | Path to the pretrained model .pth file |
--folder_lq | Input folder for low-quality images |
--folder_gt | Ground-truth folder for evaluation |
--tile | Enables tiling to manage memory (useful for real_sr) |
--large_model | Flag for using larger models in real_sr |
--noise | Noise level for denoising tasks |
--jpeg | JPEG quality factor for artifact reduction |
--large_model | Flag for using larger models in real_sr |