[Abstract]

Diffusion probabilistic models, often referred to as diffusion models, have emerged as the state-of-the-art in generative AI, but their inherently slow sampling speed remains a critical bottleneck for real-world applications. This seminar explores the fundamental trade-off between sampling efficiency and generation performance. To systematically address this challenge, we categorize the solutions into several core research scopes---training, solver enhancement, time step scheduling, and knowledge distillation---which serve as complementary approaches.

In this talk, I will introduce selected works across these perspectives. By examining these approaches, we will discuss how these distinct strategies effectively mitigate the efficiency bottleneck without compromising generation quality. Finally, the talk will conclude by briefly discussing related open challenges and future directions.

[Biography]

Jinyoung Choi is an Assistant Professor in the Graduate School of AI and the Department of CSE at UNIST. She earned her Ph.D. in Electrical and Computer Engineering from Seoul National University (SNU). Prior to her doctoral studies, she worked at LG PRI as a senior research engineer. She received her M.S. in Mathematics and B.S. in Industrial & Management Engineering and Mathematics from POSTECH.