IQA-PyTorch is a PyTorch toolbox for image quality assessment that implements both full-reference and no-reference metrics.
The toolbox addresses the need to evaluate image quality across diverse scenarios by providing PyTorch implementations of established metrics including PSNR, SSIM, LPIPS, FID, NIQE, MUSIQ, TOPIQ, NIMA, DBCNN, BRISQUE, and others. The approach reimplements these metrics in pure Python and PyTorch with results calibrated against official MATLAB implementations where available. GPU acceleration enables significantly faster computation compared to MATLAB counterparts.
Developers should choose this toolbox when they need to integrate image quality assessment into Python-based workflows, particularly those already using PyTorch. It suits projects requiring both traditional metrics like PSNR and SSIM alongside modern learned metrics and blind quality assessment methods. The toolbox provides model cards documenting each metric, dataset preparation guides, and benchmark protocols to help users select appropriate metrics for their use cases.
The project maintains active development with regular updates to its metric implementations and model weights. Documentation is comprehensive and continuously maintained through a dedicated documentation site. The toolbox is distributed as a pip-installable package with straightforward setup, and model weights are hosted on accessible mirrors to support users in regions with connectivity constraints.
