Research map / Image generation / Fast diffusion sampling
An Inversion-based Measure of Memorization for Diffusion Models
Zhe Ma, Qingming Li, Xuhong Zhang, Tianyu Du, Ruixiao Lin, Zonghui Wang, Shouling Ji, Wenzhi Chen
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 |
|---|---|---|---|
| use adaptive weight tuning of lambda instead of fixed lambda=1 (ablation) | InvMM inversion · CelebAHQ (LDM) | images with infinite memorization score: 92.4% fail at lambda=1; all inverted with adaptive | Improved |
| use adaptive lambda instead of fixed lambda=0 (ablation) | InvMM inversion · CelebAHQ, FFHQ | memorization score estimate | Improved |
| InvMM vs eps0-loss for detection | DDPM CIFAR-10 · CIFAR-10 detection | TPR@1%FPR: 0.414 -> 0.899 | Improved |
| InvMM vs Wen et al. for detection | SD v1.4 · LAION detection | AUC: 0.999 vs 0.974 | No clear effect |
| invert prompt together with noise to reveal worst-case memorization | InvMM · LAION suspicious set | memorization found | Improved |
| increase training epochs (ablation) | LDM on CelebAHQ/FFHQ · InvMM score | memorization | Hurt |
| increase data duplication (ablation) | LDM on CelebAHQ-2.5k and FFHQ-6k · InvMM score | memorization | Hurt |
| deduplicate training set (SD v2.1 and v3.5 vs v1.4) (ablation) | Stable Diffusion · LAION confirmed and suspicious sets | memorization | Improved |
| increase dataset size and complexity (ablation) | LDM on CelebAHQ/FFHQ · InvMM score | memorization | No clear effect |
| use x0-prediction objective for inversion instead of eps0-prediction (ablation) | SD v1.4 prompt inversion · confirmed set | inversion stability | Improved |
| use Gumbel-Softmax temperature 2.0 without discretization (ablation) | SD prompt inversion · LAION subsets | denoising error | Improved |
| membership inference (SecMI) vs InvMM | LDM on FFHQ-6k · FFHQ-6k | membership vs memorization | Mixed |
Closest papers by meaning
- On the Edge of Memorization in Diffusion Models (NeurIPS 2025 · similarity 0.92)
- Towards Memorization-Free Diffusion Models (CVPR 2024 · similarity 0.90)
- Does Generation Require Memorization? Creative Diffusion Models using Ambient Diffusion (ICML 2025 · similarity 0.89)
- Detecting, Explaining, and Mitigating Memorization in Diffusion Models (ICLR 2024 · similarity 0.89)
- Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization (CVPR 2025 · similarity 0.89)
- Understanding and Mitigating Memorization in Generative Models via Sharpness of Probability Landscapes (ICML 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.
