Research map / Probabilistic and scientific modeling / Variational and generative sampling
Effectively Unbiased FID and Inception Score and Where to Find Them
Check 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 |
|---|---|---|---|
| Use Sobol sampling instead of IID normal sampling for finite-sample metric evaluation | GAN evaluation · CelebA, DCGAN | FID (50k): 14.61 $\pm$ 0.0579 -> 14.58 $\pm$ 0.0439 | Helped |
| Use Sobol sampling instead of IID normal sampling for finite-sample metric evaluation | BigGAN · ImageNet | IS (5k): 92.96 $\pm$ 2.135 -> 93.21 $\pm$ 1.640 | Helped |
| Use Sobol sampling instead of IID normal sampling for finite-sample metric evaluation | BigGAN · CIFAR10 | IS (5k): 8.44 $\pm$ 0.1223 -> 8.48 $\pm$ 0.1172 | Helped |
| Use Sobol sampling with inverse CDF rather than Box-Muller (ablation) | GAN metric evaluation · ImageNet, BigGAN | IS (5k): 92.89 $\pm$ 1.961 -> 93.21 $\pm$ 1.640 | Helped |
| Use Sobol sampling with inverse CDF rather than Box-Muller (ablation) | GAN metric evaluation · CIFAR10, BigGAN | IS (5k): 8.48 $\pm$ 0.1172 -> 8.45 $\pm$ 0.1018 | Mixed |
| Use Sobol sampling with inverse CDF rather than Box-Muller (ablation) | GAN metric evaluation · CelebA, DCGAN | FID (50k): 14.59 $\pm$ 0.0471 -> 14.58 $\pm$ 0.0439 | Helped |
| Use Sobol sampling with inverse CDF rather than Box-Muller (ablation) | GAN metric evaluation · CelebA, ProGAN | FID (50k): 9.94 $\pm$ 0.0384 -> 9.94 $\pm$ 0.0404 | No clear effect |
| Use Sobol sampling with inverse CDF rather than Box-Muller (ablation) | GAN metric evaluation · Flickr-Faces-HQ, StyleGAN | FID (50k): 4.33 $\pm$ 0.0406 -> 4.33 $\pm$ 0.0354 | Helped |
| Extrapolate finite-sample FID values in 1/N to estimate the infinite-sample score | FID evaluation · Generative models | FID bias | Helped |
| Evaluate at a fixed finite sample count rather than extrapolating | FID and IS comparisons · Cross-generator comparisons | comparison reliability | Hurt |
| Choose evaluation sample sizes at regular intervals over N rather than over 1/N (ablation) | FID extrapolation · FID100k estimation | extrapolation accuracy | Helped |
| Use Sobol_Inv rather than normal sampling when fitting the FID extrapolation (ablation) | FID extrapolation · FID100k estimation | estimate accuracy and variance | Helped |
Closest papers by meaning
- Rethinking FID: Towards a Better Evaluation Metric for Image Generation (CVPR 2024 · similarity 0.90)
- The Role of ImageNet Classes in Fréchet Inception Distance (ICLR 2023 · similarity 0.89)
- Assessing Generative Models via Precision and Recall (NeurIPS 2018 · similarity 0.88)
- MIND: Monge Inception Distance for Generative Models Evaluation (NeurIPS 2026 · similarity 0.87)
- Representation Fréchet Loss for Visual Generation (NeurIPS 2026 · similarity 0.85)
- Anomaly Score: Evaluating Generative Models and Individual Generated Images based on Complexity and Vulnerability (CVPR 2024 · similarity 0.83)
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.