Research map / Image generation / Fast diffusion sampling
Detecting and Mitigating Memorization in Diffusion Models through Anisotropy of the Log-Probability
Rohan Asthana, Vasileios Belagiannis
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 anisotropic cosine-similarity term to norm-based metric (ablation) | norm of score difference metric · Detection SD v1.4 (n=1) | TPR@1%FPR: 0.896 -> 0.934 | Improved |
| add anisotropic cosine-similarity term to norm-based metric (ablation) | norm of score difference metric · Detection SD v2.0 (n=1) | AUC: 0.948 -> 0.952 | Improved |
| add isotropic norm term to cosine-similarity metric (ablation) | cosine similarity in anisotropy · Detection SD v1.4 (n=1) | TPR@1%FPR: 0.424 -> 0.934 | Improved |
| use fitted gamma weights instead of equal weights (ablation) | metric M · Detection SD v2.0 (n=4) | TPR@1%FPR: 0.868 -> 0.900 | Improved |
| swap cosine target to conditional score vs unconditional score (ablation) | metric M · Detection SD v1.4 (n=1) | AUC: 0.980 vs 0.992 | No clear effect |
| use M instead of best prior denoising-free method | sharpness-based detector · Detection SD v1.4 (n=1) | TPR@1%FPR: 0.908 -> 0.935 | Improved |
| use M instead of best prior denoising-free method | sharpness-based detector · Detection SD v2.0 (n=4) | AUC: 0.991 -> 0.981 | Hurt |
| use M instead of best prior denoising-free method | sharpness-based detector · Detection speed SD v1.4 (n=1) | time (s) per 10 prompts: 5.40 -> 1.10 | Improved |
| use M instead of best prior denoising-free method | sharpness-based detector · Detection SD v2.0 (n=1) | TPR@1%FPR: 0.740 -> 0.791 | Improved |
| evaluate norm metric in anisotropic (t~0) vs isotropic (t~T) regime (ablation) | Wen norm-based metric · memorized vs non-memorized separation | KL divergence: 0.166 -> 0.022 | Hurt |
Closest papers by meaning
- Understanding and Mitigating Memorization in Generative Models via Sharpness of Probability Landscapes (ICML 2025 · similarity 0.91)
- Towards Memorization-Free Diffusion Models (CVPR 2024 · similarity 0.90)
- Detecting, Explaining, and Mitigating Memorization in Diffusion Models (ICLR 2024 · similarity 0.90)
- How Diffusion Models Memorize (NeurIPS 2026 · similarity 0.90)
- Cyclic Denoising Reveals Ultrastable Memories in Diffusion Models (NeurIPS 2026 · similarity 0.89)
- Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.89)
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.
