CSE Colloquium: Accelerated Diffusion Models

Presenter: Julius Berner, NVIDIA

Abstract: Diffusion and flow models have driven major advances in image, video, and 3D content generation, as well as in a wide range of scientific applications. As the field moves toward interactive applications—including real-time editing, world models, and embodied AI—low-latency generation is becoming essential. This talk will begin with an overview of techniques for accelerating diffusion and flow models, then examine step-distillation methods in detail. We will cover both trajectory-based approaches, such as knowledge distillation, consistency models, and flow maps, and distribution-based approaches, such as adversarial distillation and variational score distillation. We will illustrate these methods with practical examples using our open-source FastGen library.

Bio: Julius Berner is a Research Scientist at NVIDIA, where he works on post-training of diffusion and flow models for real-time media generation and interactive world models. His broader research spans generative modeling, deep learning, and scientific applications. Prior to NVIDIA, he was a postdoc at Caltech and completed his PhD sub auspiciis praesidentis at the University of Vienna, alongside research internships at Meta and NVIDIA.

Hosted By: Professor Daniel Fremont and Professor Leilani Gilpin

Zoom: https://ucsc.zoom.us/j/95300285113?pwd=62tnEPUtRrsmbW7SHTavaRbaF2afSr.1