Research map / Image generation / Fast diffusion sampling
Provable Separations between Memorization and Generalization in Diffusion Models
Zeqi Ye, Qijie Zhu, Molei Tao, Minshuo 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 |
|---|---|---|---|
| add importance-based attention-head pruning (20%) plus fine-tuning | trained DiT · CIFAR-10 (5000-sample subset) | memorization ratio (%): 73.82 -> 68.58 | Improved |
| add importance-based head pruning | trained DiT · CIFAR-10 (5000-sample subset) | recall: 0.08 -> 0.12 | Improved |
| add importance-based head pruning | trained DiT · CIFAR-10 (5000-sample subset) | FID: 15.47 -> 15.07 | Improved |
| add importance-based head pruning | trained DiT · CIFAR-10 (5000-sample subset) | precision: 0.39 -> 0.33 | Hurt |
| use importance-based pruning instead of random pruning (ablation) | pruned DiT · CIFAR-10 (5000-sample subset) | FID: 17.14 -> 15.07 | Improved |
| use importance-based pruning instead of random pruning (ablation) | pruned DiT · CIFAR-10 (5000-sample subset) | recall: 0.09 -> 0.12 | Improved |
| use importance-based pruning instead of random pruning (ablation) | pruned DiT · CIFAR-10 (5000-sample subset) | memorization ratio (%): 66.87 -> 68.58 | Hurt |
| increase training sample size (ablation) | diffusion model on Gaussian mixture · Gaussian mixture (d=32) | memorization ratio | Improved |
| increase data dimension (ablation) | diffusion model on Gaussian mixture · Gaussian mixture, 12M-parameter network | memorization ratio | Improved |
| add strong weight decay with ample data (ablation) | diffusion model on Gaussian mixture · Gaussian mixture (n=10K) | mean log-likelihood | Hurt |
| add proper weight decay and moderate width with few samples (ablation) | diffusion model on Gaussian mixture · Gaussian mixture (n=3.2K) | memorization ratio | Improved |
Closest papers by meaning
- On the Edge of Memorization in Diffusion Models (NeurIPS 2025 · similarity 0.94)
- Understanding diffusion models requires rethinking (again) generalization (NeurIPS 2026 · similarity 0.92)
- On the Memorization of Consistency Distillation for Diffusion Models (NeurIPS 2026 · similarity 0.91)
- Detecting, Explaining, and Mitigating Memorization in Diffusion Models (ICLR 2024 · similarity 0.91)
- How Diffusion Models Memorize (NeurIPS 2026 · similarity 0.91)
- Does Generation Require Memorization? Creative Diffusion Models using Ambient Diffusion (ICML 2025 · similarity 0.91)
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.
