Research map / Probabilistic and scientific modeling / Variational and generative sampling
Enhanced Generative Model Evaluation with Clipped Density and Coverage
Nicolas Salvy, Hugues Talbot, Bertrand Thirion
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| clip real ball radii to the median k-NN distance (ablation) | Density · CIFAR-10 RESFLOW | Density: 2.47 -> 0.00 | Improved |
| clip real ball radii (ablation) | Density · CIFAR-10 NVAE | Density: 1.59 -> 0.00 | Improved |
| clip real ball radii (ablation) | Density · ImageNet DiT-XL-2-guided | over-occurring samples (%): 63.8 -> 42.3 | Improved |
| clip real ball radii (ablation) | Density · CIFAR-10 LSGM-ODE | over-occurring samples (%): 18.4 -> 6.0 | Improved |
| clip individual sample contributions and normalize (ablation) | Density · CIFAR-10 LSGM-ODE | Density: 0.22 -> 0.14 | Improved |
| clip radii of real balls (ablation) | Density · CIFAR-10 RESFLOW | real balls with no synthetic point: 48373 -> 49981 | Improved |
| replace pairwise distance matrices with scikit-learn nearest neighbors | Precision Cover · DINOv2-FFHQ-LDM | time (min): 88.2 -> 2.90 | Improved |
| replace pairwise distance matrices with scikit-learn nearest neighbors | Recall Cover · Gaussians d=32 | time (s): 61.0 -> 2.25 | Improved |
| replace pairwise distance matrices with scikit-learn nearest neighbors | P-precision / P-recall · DINOv2-FFHQ-LDM | time (min): DNF -> 0.7229 | Improved |
| replace pairwise distance matrices with nearest neighbors | Precision, Recall, Density, Coverage · 10000 Gaussians d=32 | time (s): 3.93 -> 8.43 | Hurt |
| vary k in Clipped Density and Coverage (ablation) | Clipped Density/Coverage · generated datasets | score | No clear effect |
Closest papers by meaning
- Position: All Current Generative Fidelity and Diversity Metrics are Flawed (ICML 2025 · similarity 0.90)
- RA-ClipScore: Making Generative Model Evaluation More Interpretable (NeurIPS 2026 · similarity 0.88)
- A Unifying Information-theoretic Perspective on Evaluating Generative Models (AAAI 2025 · similarity 0.88)
- Attribute Based Interpretable Evaluation Metrics for Generative Models (ICML 2024 · similarity 0.87)
- PQMass: Probabilistic Assessment of the Quality of Generative Models using Probability Mass Estimation (ICLR 2025 · similarity 0.86)
- On the Stability of Iterative Retraining of Generative Models on their own Data (ICLR 2024 · similarity 0.86)
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.
