Kynkäänniemi, Karras, Aittala, Aila, Lehtinen, Aalto and NVIDIA (2022; ICLR 2023) — ARXIV-2203.06026
Key takeaways
- What FID looks at. A Grad-CAM (LIT-732) adaptation shows which image regions move FID. On FFHQ it is mostly outside the face, wherever the Top-1 ImageNet class (bow tie, seat belt, mortarboard) sits. ImageNet has no “person” or “face” class
- The perceptual null space is large. With the generator fixed:
- Top-1 histogram matching by rejection sampling lowers FID by about 10% on five datasets. CLIP-space FD moves about 1%
- Optimizing sample weights over 5× oversampled candidates to match all “fringe features” lowers FFHQ FID 5.30 → 1.78 (−66%). ResNet-50 FD falls 37%, SwAV-FD 13% and CLIP-FD 4%. The drop tracks how much ImageNet each feature space has seen
- A practical case (§4): Projected FastGAN (LIT-562), trained with its public code on FFHQ, reaches FID 5.28 against StyleGAN2’s 5.30 at similar recall. CLIP-FD is 4.67 against 2.76, and human raters prefer StyleGAN2. The paper says Sauer et al.’s own human study agreed
- Where FID stays reliable, per the authors: detecting failed runs, monitoring one run’s convergence, and comparing hyperparameter sweeps within one setup. Where it breaks: comparing sufficiently different architectures or training setups, especially where one uses ImageNet pretraining
Standing in the anthology
Filed from #163 while working on Projected GAN, because it is the
measurement that contests it. It sources SOTA-337 and is the
contested_by of SOTA-338.
A caveat on the source. The authors include StyleGAN’s authors, and the paper’s worked example is a competitor losing to StyleGAN2. The resampling experiment, which involves no competitor, is the evidence that does not depend on that comparison.
It sits beside the record’s other FID practices, SOTA-307 (report FID over seeds) and the tokenizer-rate practice. Those are about FID’s variance and comparability. This one is about its validity.
What it takes from FID’s definition. It keeps the metric of Heusel et al. (LIT-611) as defined, a Fréchet distance between Gaussians fitted to Inception-V3 pool3 features, and asks what that feature space responds to, which the original never examined: its validation was only that FID rises monotonically as images are deliberately corrupted. The addition is the attribution analysis and the resampling experiments, plus the same distance computed in ResNet-50, SwAV and CLIP feature spaces to show how much of a gain depends on Inception’s ImageNet training.
Read — NOTE-304.