Parham Rezaei

Through controlled experiments, we found that AR models trained on coarse-to-fine ordered tokens exhibit improved test-time scaling behavior compared with grid-based counterparts.

We analyzed data augmentation with generative models from a random matrix theory perspective.

We proposed a probabilistic view of spatial relationship alignment in text-to-image models and used it to improve how well generated images follow textual descriptions.

We proposed an online algorithm for selecting diverse mixtures of generative models.

We investigated diffusion models in text-to-image generation and found that they outperform autoregressive models on compositional generation tasks.