Parham Rezaei

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

We analyzed augmenting data from a generative model from the perspective of random matrix theory.

We proposed a new way to look at the spatial relationship alignment in T2I models and how to use it to improve the adherence of generated images to textual descriptions.

We proposed an online algorithm for finding the optimal mixture of generative models.

We investigated the performance of diffusion models in text-to-image generation tasks. We found that diffusion models outperform autoregressive models in terms of compositional generation.