SDS Seminar Series – Mingyuan Zhou, University of Texas at Austin

Oct

2024

2:00 pm – 3:00 pm

In Person (view details)

Featured Speaker(s): Mingyuan Zhou

Cost: Free

Building Faster, Better, and Safer Deep Generative Models via Score Identity Distillation

Description

The Fall 2024 SDS Seminar Series continues on October 11th from 2:00 p.m. to 3:00 p.m. with Dr. Mingyuan Zhou (Associate Professor, Departments of Information, Risk & Operations Management and Statistics and Data Sciences, University of Texas at Austin). This event is in-person in CBA 4.348.

Title: Building Faster, Better, and Safer Deep Generative Models via Score identity Distillation

Abstract: Diffusion models, celebrated for generating photorealistic outputs, face significant challenges: slow generation times and the potential to produce inappropriate content. Our Score identity Distillation (SiD) method addresses these issues head-on. Contrary to the prevailing belief that high-quality diffusion models require iterative refinement, SiD introduces a revolutionary single-step generative process. This approach not only speeds up generation but, in many instances, surpasses the quality of original models, which typically require extensive iterative steps—from dozens to hundreds. Further enhancements are achieved through our innovative method of reintroducing training data to perform joint distillation and adversarial generation, significantly extending performance boundaries. Moreover, we adapt SiD to selectively forget unsafe concepts such as nudity and personal identities, enhancing the safety of the generated content. These advancements make high-quality generative processes more accessible and practical for a broad range of applications, paving the way for new research and practical uses in generative AI.