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 †‡ · 2026
Read Two Calm Ends and the Wild Middle: A Geometric Picture of Memorization in Diffusion Models by Nick Dodson ∗, Xinyu Gao ∗, Qingsong Wang ∗, Yusu Wang †, Zhengchao Wan †‡ (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-17846.txt.
Description
Nick Dodson and team give a geometric picture of memorization in diffusion models. Diffusion models generate high-quality samples but can memorize training data, raising privacy concerns. It remains unclear where along the noise schedule memorization arises. It is also unclear how data geometry shapes it, and how effects at varied noise scales interact.
The authors partition the noise schedule into three regimes. The split rests on coverage of training data by Gaussian shells and concentration of the posterior, two objects argued to govern memorization and generalization. Memorization risk proves highly non-uniform across noise levels, with a danger zone at medium noise where memorization peaks. Small and large noise regimes resist memorization through distinct mechanisms: small noise through limited training coverage, large noise through low posterior concentration and provably near-linear Gaussian denoising. For medium noise, geometric conditions support a targeted, geometry-informed intervention that mitigates memorization.
Catalog record
| CREATOR | Nick Dodson ∗, Xinyu Gao ∗, Qingsong Wang ∗, Yusu Wang †, Zhengchao Wan †‡ |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 95 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-17846