Research map / Probabilistic and scientific modeling / Molecules and proteins
GICDM: Mitigating Hubness for Reliable Distance-Based Generative Model Evaluation
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 |
|---|---|---|---|
| add GICDM hubness correction | Clipped Density · Synthetic benchmark (Purpose) | passed tests out of 14: 8/14 -> 10/14 | Improved |
| add GICDM hubness correction | Clipped Density · Synthetic benchmark (Bounds) | passed tests out of 13: 8/13 -> 11/13 | Improved |
| add GICDM hubness correction | Clipped Coverage · Synthetic benchmark (Purpose) | passed tests out of 14: 8/14 -> 10/14 | Improved |
| add GICDM hubness correction | Clipped Coverage · Synthetic benchmark (Bounds) | passed tests out of 13: 9/13 -> 11/13 | Improved |
| add GICDM hubness correction | Clipped Density · Synthetic benchmark (all criteria) | passed criteria: 18 -> 23 | Improved |
| add GICDM hubness correction | Clipped Coverage · Synthetic benchmark (all criteria) | passed criteria: 18 -> 22 | Improved |
| add GICDM hubness correction | Clipped Density (DINOv3) · CIFAR-10 human correlation | Pearson correlation: 0.82 -> 0.95 | Improved |
| add GICDM hubness correction | Clipped Density (DINOv3) · ImageNet human correlation | Pearson correlation: 0.67 -> 0.82 | Improved |
| add GICDM hubness correction | Clipped Density (DINOv2) · CIFAR-10 human correlation | Pearson correlation: 0.93 -> 0.97 | Improved |
| add GICDM hubness correction | Clipped Density (DINOv2) · LSUN Bedroom human correlation | Pearson correlation: 0.81 -> 0.80 | No clear effect |
| add GICDM hubness correction | Clipped Coverage (DiT-XL-2, DINOv2) · ImageNet classifier-free guidance | Clipped Coverage with guidance: 0.64 -> 0.83 without GICDM; 0.79 -> 0.55 with GICDM | Improved |
| replace NICDM with iterative ICDM (ablation) | hubness reduction · 5 datasets/embeddings avg | h_1^5(1%): 3.1 -> 1.7 | Improved |
Closest papers by meaning
- Efficient Precision and Recall Metrics for Assessing Generative Models using Hubness-aware Sampling (ICML 2024 · similarity 0.90)
- The Fault in Our Metrics: Revisiting Generative Model Evaluation with Chamfer Distance (NeurIPS 2026 · similarity 0.87)
- A Theoretical Framework for Statistical Evaluability of Generative Models (ICML 2026 · similarity 0.87)
- RA-ClipScore: Making Generative Model Evaluation More Interpretable (NeurIPS 2026 · similarity 0.86)
- A Unifying Information-theoretic Perspective on Evaluating Generative Models (AAAI 2025 · similarity 0.86)
- Distance Marching for Generative Modeling (NeurIPS 2026 · 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.
