Research map / Image generation / Autoregressive and latent generation
Reliable Fidelity and Diversity Metrics for Generative Models
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 |
|---|---|---|---|
| Replace improved precision with density | Generative-model evaluation on matching real and fake distributions · 64-dimensional Gaussian | fidelity score: 0.68 -> 1.06 | Helped |
| Replace improved recall with coverage | Generative-model evaluation on matching real and fake distributions · 64-dimensional Gaussian | diversity score: 0.67 -> 0.97 | Helped |
| Add density-based counting of overlapping real-sample neighborhoods instead of binary precision membership | Density and coverage evaluation · Real or fake samples with one outlier | fidelity robustness | Helped |
| Build coverage neighborhoods around real samples rather than fake samples | Density and coverage evaluation · CelebA and LSUN-bedroom fake-sample outliers | coverage: less than 2% increase | Helped |
| Build recall neighborhoods around fake samples | Improved precision and recall · CelebA and LSUN-bedroom fake-sample outliers | recall: more than 11% and 15% | Hurt |
| Use coverage instead of recall to measure diversity | Generative-model evaluation · Simultaneous mode dropping in Gaussian mixtures | diversity sensitivity | Helped |
| Use coverage instead of recall to measure diversity | Generative-model evaluation · MNIST class mode dropping | diversity sensitivity | Helped |
| Truncate the latent variable at magnitude 1 (ablation) | Convolutional VAE · MNIST | coverage: 0.411 -> 0.291 | Hurt |
| Truncate the latent variable at magnitude 1 (ablation) | Convolutional VAE · MNIST | density: 0.208 -> 0.205 | No clear effect |
| Use randomly initialized VGG16 features instead of ImageNet-pretrained features | Generative-model evaluation · MNIST and SC09 sound | evaluation scores: 0.047 density -> 0.491 density; 0.029 recall -> 0.572; 0.020 coverage -> 0.653 | Helped |
Closest papers by meaning
- A Unifying Information-theoretic Perspective on Evaluating Generative Models (AAAI 2025 · similarity 0.90)
- Rethinking FID: Towards a Better Evaluation Metric for Image Generation (CVPR 2024 · similarity 0.89)
- Position: All Current Generative Fidelity and Diversity Metrics are Flawed (ICML 2025 · similarity 0.89)
- Efficient Precision and Recall Metrics for Assessing Generative Models using Hubness-aware Sampling (ICML 2024 · similarity 0.88)
- Image Generation Diversity Issues and How to Tame Them (CVPR 2025 · similarity 0.87)
- Anomaly Score: Evaluating Generative Models and Individual Generated Images based on Complexity and Vulnerability (CVPR 2024 · similarity 0.87)
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.