chaofengc/IQA-PyTorch

What it solves

PyIQA is a comprehensive toolbox for Image Quality Assessment (IQA), providing a unified way to measure how "clean" or high-quality an image is. It solves the problem of fragmented implementations of quality metrics, which were often written in MATLAB and were slow. By reimplementing these in pure Python and PyTorch, the project provides GPU acceleration and a standardized interface for both full-reference (FR) and no-reference (NR) metrics.

How it works

The toolbox provides a library of pre-implemented metrics (such as LPIPS, SSIM, and FID) and a command-line interface for quick testing. It allows users to create metrics via pyiqa.create_metric, which can be used for simple score inference or as loss functions in training pipelines (with as_loss=True). It also includes built-in support for loading popular IQA datasets and provides scripts for benchmarking performance across different datasets.

Who it’s for

Researchers and developers working in computer vision, image restoration, and generative AI who need to objectively measure image quality, evaluate generative models, or use quality metrics as differentiable loss functions during model training.

Highlights

• GPU Accelerated: Significantly faster than traditional MATLAB implementations.
• Comprehensive Metric Zoo: Supports a wide range of full-reference, no-reference, aesthetic, and face-specific IQA metrics.
• Differentiable: Many metrics can be used as loss functions for backpropagation.
• Dataset Integration: Simplified loading and downloading of popular IQA datasets via Hugging Face.
• Extensible: Experimental support for plugin authoring to add custom datasets, architectures, and metrics.