clean-fid

Why we included this project

FID scores only mean something when the resizing and quantization steps behind them match what the reference implementations did, and small differences there can shift the number by more than the gap between competing models. clean-fid is a compact PyTorch library that gets those steps right, so the FID you compute lines up with the values quoted in papers instead of drifting. The compute_fid and compute_kid calls work against image folders, precomputed dataset statistics such as FFHQ and CIFAR-10, or a generator function directly, and CLIP-feature FID is supported too. For anyone training image generators, it is the practical way to benchmark against published baselines and tell whether a reported improvement is real.