GaParmar/clean-fid

Standardizes FID computation by addressing critical implementation inconsistencies in image resizing (aliasing artifacts across libraries) and JPEG compression effects that can inflate scores by 6+ points. Provides pre-computed statistics for standard benchmarks (CIFAR-10, FFHQ, LSUN) and supports alternative metrics like KID and CLIP-FID features via unified Python APIs. Enables direct evaluation against generative models or image folder pairs with configurable resizing filters and quantization to ensure reproducible, comparable results across research groups.

1,140 stars. No commits in the last 6 months. Available on PyPI.