Image quality metrics that train with your model

What it does

PIQ is a PyTorch-native collection of image quality metrics spanning full-reference (PSNR, SSIM, LPIPS), no-reference (BRISQUE, CLIP-IQA), and distribution-based (FID, KID, Inception Score) measures. It exposes both functional interfaces for quick evaluation and class-based modules that plug directly into training loops as loss functions. The library targets generative and image-to-image tasks where you need consistent, validated scoring without framework hopping.

The interesting bit

Most metric implementations stop at evaluation; PIQ makes the majority of them fully differentiable so you can backpropagate through SSIM, FID, or MS-SSIM during model optimization. It also ships with benchmarking scripts against standard opinion-score datasets like TID2013 and KADID10k, letting you verify that its SSIM actually matches the literature’s SSIM.

Key highlights

• Pure PyTorch with minimal dependencies; runs on GPU and stays inside your existing training graph.
• Covers three metric families: full-reference, no-reference, and distribution-based.
• Most metrics implement both functional and module interfaces, so the same code handles evaluation and loss computation.
• Includes validation logic to catch bad inputs before they crash a long training run.
• Benchmarks against TID2013, KADID10k, PIPAL, KonIQ10k, and LIVE-itW with published SRCC comparisons.

Caveats

• The README notes that compute_feats for distribution metrics expects a data loader of “predefined format,” but does not specify what that format is.
• Python support is explicitly capped at 3.7–3.10.
• Benchmark datasets are not bundled; you must download TID2013, KADID10k, and others separately.

Verdict

Worth a look if you train GANs, autoencoders, or any image-to-image models in PyTorch and are tired of copy-pasting metric implementations. Skip it if you only need a single static metric and do not care about differentiability or unified APIs.