Generalization, memorization, and overfitting for diffusion models trained in the lazy high-dimensional regime
stat.ML, cs.LG, math.ST, stat.TH
Submitted: 2026-08-25
Updated: 2026-08-25
Comments: 100 pages, 3 figures
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Terminology
Sources
- Losing dimensions: Geometric memorization in generative diffusion
- Memorization and Regularization in Generative Diffusion Models
- Anti-Concentration Inequalities for the Difference of Maxima of Gaussian Random Vectors
- High-accuracy sampling for diffusion models and log-concave distributions
- Quantitative deterministic equivalent of sample covariance matrices with a general dependence structure
- From optimal score matching to optimal sampling
- Benign Overfitting Does Not Occur in Diffusion Models
- Anisotropic local law for non-separable sample covariance matrices
- Bigger Isn't Always Memorizing: Early Stopping Overparameterized Diffusion Models
- Precise Performance of Linear Denoisers in the Proportional Regime
- Asymptotic Learning Curves for Diffusion Models with Random Features Score and Manifold Data
- An analytic theory of creativity in convolutional diffusion models
- Diffusion Models Memorize in Training -- and Generalize in Inference
- Generalization Dynamics of Linear Diffusion Models
- Local Coverage Governs Memorization in Diffusion Models
- A non-asymptotic theory of Kernel Ridge Regression: deterministic equivalents, test error, and GCV estimator
- Universality laws for random matrices via exchangeable counterparts
- The Hidden Linear Structure in Score-Based Models and its Application
- A Random Matrix Theory Perspective on the Consistency of Diffusion Models
Related papers
- Behavior of prediction performance metrics with rare events
- Optimal Estimation of Generic Dynamics by Path-Dependent Neural Jump ODEs
- A Posterior-Dynamics Framework for Imaging Inverse Problems with Pretrained Diffusion Priors
- One Permutation Is All You Need: Fast, Deterministic Feature Importance and Model Stress-Testing
- Online Conformal Prediction for Non-Exchangeable Panel Data
- Deep Time-Series Forecasting in 10 Years: A Survey