Selected Publications
* indicates equal contribution.
Scaling and Distilling Text Embeddings for Better Diffusibility
We distill scaled text embeddings into a good latent for continuous diffusion language models.
Generalization of Diffusion Models Arises with a Balanced Representation Space
Learning good representations is central to novel and meaningful generative modelling.
A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective
Diffusion models collapse to a low-diversity/quality mode when trained on their own outputs.
Understanding Representation Dynamics of Diffusion Models via Low-Dimensional Modeling
How time conditioning shapes diffusion representations and how it can diagnose overfitting.