Please note: This PhD seminar will take place online.
Haoye Lu, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Yaoliang Yu
In recent years, diffusion models have risen to prominence as the foremost technique for distribution learning. The related paradigms also include flow matching, rectified flow and bridge models.
In this talk, I will begin by examining the close relationships among these paradigms, demonstrating how they can be unified within a universal framework. Next, I will discuss the reflow process, designed to straighten the flow trajectory used in rectified flow and its variants, eliminating the need for iterative training across multiple flow models. Finally, I will discuss a special family of flow models that harness the intrinsic symmetries within data distributions. This approach enables theoretically robust equivariant image noise reduction and style transfer, regardless of the image’s orientation — a crucial factor in medical image analysis to achieve consistent diagnostic outcomes.