Research map / Image generation / Autoregressive and latent generation
The Role of ImageNet Classes in Fréchet Inception Distance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Match the Top-1 ImageNet class histogram by resampling generated images | StyleGAN2 · FFHQ and other tested datasets | FID: ∼10% | Helped |
| Match the Inception-V3 pre-logit distribution by optimizing image sampling weights | StyleGAN2 · Tested datasets | FID: as much as 60% | Helped |
| Optimize pre-logit sampling weights | StyleGAN2 · FFHQ | FID: -66.4% | Helped |
| Match Top-N ImageNet class co-occurrence by optimizing sampling weights | StyleGAN2 · FFHQ | FID | Helped |
| Align Top-N class histograms | StyleGAN2 · LSUN Cat, LSUN Car, LSUN Places, and AFHQ-v2 Dog | FID | Helped |
| Use Inception-V3 rather than non-ImageNet-trained feature spaces to assess the resampling improvement (ablation) | Top-1 histogram-matched StyleGAN2 samples · Tested datasets | FID: ∼1% | Mixed |
| Optimize resampling weights in pre-logit space rather than logit space (ablation) | StyleGAN2 · Tested datasets | FID | No clear effect |
| Match middle-ranked ImageNet classes instead of the top-ranked classes (ablation) | StyleGAN2 · FFHQ | FID | Hurt |
| Use CLIP features to evaluate ImageNet-class resampling rather than Inception-V3 features (ablation) | Resampled StyleGAN2 · FFHQ | FID | No clear effect |
| Use ImageNet-pretrained discriminator features in Projected FastGAN | Projected FastGAN · FFHQ | human-rated image quality | Hurt |
| Compare Projected FastGAN and StyleGAN2 using CLIP FID rather than Inception FID | Projected FastGAN and StyleGAN2 · FFHQ | FID_CLIP | Mixed |
| Add Gaussian noise to FID-important regions rather than equally sized unimportant regions | Generated images · FID sensitivity validation | FID | Hurt |
Closest papers by meaning
- Rethinking FID: Towards a Better Evaluation Metric for Image Generation (CVPR 2024 · similarity 0.92)
- Representation Fréchet Loss for Visual Generation (NeurIPS 2026 · similarity 0.89)
- MIND: Monge Inception Distance for Generative Models Evaluation (NeurIPS 2026 · similarity 0.88)
- Fréchet Wavelet Distance: A Domain-Agnostic Metric for Image Generation (ICLR 2025 · similarity 0.86)
- ImageNet-D: Benchmarking Neural Network Robustness on Diffusion Synthetic Object (CVPR 2024 · similarity 0.84)
- STREAM: Spatio-TempoRal Evaluation and Analysis Metric for Video Generative Models (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.