主讲人
Jie Zhu
Peking University
时间
2026年4月14日 星期二
下午 14:00-15:00
地点
学院104会议室
Abstract
Diffusion model has attracted vast attention due to its impressive image generation capabilities. However, existing models still face challenges in generating natural results, achieving efficient training, and ensuring safety. This work presents a systematic study of key technologies for diffusion models. Specifically, we focus on 1) generation quality, 2) training efficiency, 3) multimodal unification, and 4) deployment safety. For example, we propose a Mixture of Low-rank Experts (MoLE) framework, accompanied by a large-scale high-quality dataset, significantly improving the realism of facial and hand details. Overall, these efforts collectively advances diffusion models toward higher quality, efficiency, generality, and trustworthiness, contributing to the development of next-generation multimodal foundation models that are both powerful and privacy-aware.
Biography
Jie Zhu currently is a fifth-year Ph.D. candidate in the School of Computer Science, Peking University, supervised by tenured Associate Professor Leye Wang. Prior to that, he obtained his bachelor degree in Beihang University in 2021. His research interests mainly involve AI security, e.g., membership inference and data privacy, and Multi-modality large language model (MLLM), e.g., visual perception, generative models, and their unification, etc. He has published six first-author papers (5 CCF-A) including CCS, ICLR, NeurIPS, ASE, IEEE TSE, and TMLR.
