Selected Publications
* indicates equal contribution.
Generalization of Diffusion Models Arises with a Balanced Representation Space
Diffusion models generalize by extracting underlying structures within the data, learning balanced and informative representations.
A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective
When diffusion models are iteratively trained with synthetic data, the generated distribution collapses toward a stable but low-diversity, low-quality mode.
Understanding Representation Dynamics of Diffusion Models via Low-Dimensional Modeling
Analyzes how time conditioning shapes diffusion representations and how these dynamics can diagnose overfitting.