Research map / Image generation / Fast diffusion sampling
Two Calm Ends and the Wild Middle: A Geometric Picture of Memorization in Diffusion Models
Nick Dodson, Xinyu Gao, Qingsong Wang, Yusu Wang, Zhengchao Wan
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 |
|---|---|---|---|
| exclude the medium-noise danger zone sigma in [1.0, 5.0] from training (ablation) | EDM on CIFAR-10 car subset · CIFAR-10 car subset (2k images) | memorization ratio (%): 78.8% -> 0.7% | Improved |
| exclude the medium-noise danger zone sigma in [1.0, 5.0] from training (ablation) | EDM on CIFAR-10 car subset · CIFAR-10 car subset (2k images) | FID: 3.72 -> 2.35 | Improved |
| add a random-noise second class (2k-Dummy) | EDM on CIFAR-10 car subset · CIFAR-10 car subset (2k images) | memorization ratio (%): 78.8% -> 9.0% | Improved |
| swap in the EDM-50K denoiser over the medium noise range (ablation) | EDM-1K · CIFAR-10 (denoiser swapping) | trajectory memorization rate: 92.2% -> 0.0% | Improved |
| swap in the EDM-50K denoiser over the large or small noise range (ablation) | EDM-1K · CIFAR-10 (denoiser swapping) | trajectory memorization rate: 92.2% -> 93.0% (large), 91.0% (small) | No clear effect |
| swap in the EDM-1K denoiser over the medium noise range (ablation) | EDM-50K · CIFAR-10 (denoiser swapping) | trajectory memorization rate: 0% -> 92.2% | Hurt |
| skip timesteps 100-350 during training (ablation) | DDPM on grayscale CelebA · Grayscale CelebA (1024 samples) | memorization (%): 75% -> 9% | Improved |
| narrow the skipped gap to timesteps 100-200 (ablation) | DDPM on grayscale CelebA · Grayscale CelebA (1024 samples) | memorization (%): 75% -> 43% | Mixed |
| cap training noise levels (sigma < 5, 3, 1) (ablation) | Stable Diffusion v1.4 finetuning · SD v1.4 finetuning (200 memorized pairs) | mean SSCD: 0.42 -> 0.29, 0.22, 0.10 | Improved |
| repeat denoising for four ODE steps from sigma = 3.3 | EDM-1K · CIFAR-10 (ODE trajectory windows) | memorization rate: 14.1% -> 91% | Hurt |
Closest papers by meaning
- On the Edge of Memorization in Diffusion Models (NeurIPS 2025 · similarity 0.92)
- How Diffusion Models Memorize (NeurIPS 2026 · similarity 0.91)
- Why Diffusion Models Don’t Memorize: The Role of Implicit Dynamical Regularization in Training (NeurIPS 2025 · similarity 0.90)
- Understanding and Mitigating Memorization in Generative Models via Sharpness of Probability Landscapes (ICML 2025 · similarity 0.90)
- Understanding diffusion models requires rethinking (again) generalization (NeurIPS 2026 · similarity 0.90)
- Cyclic Denoising Reveals Ultrastable Memories in Diffusion Models (NeurIPS 2026 · 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.
