일시: 2025. 7. 8 (화요일), 오후 4시

장소: 제4공학관, D504

Speaker: Jun-Yan Zhu, Ph.D. (Michael B. Donohue Assistant Professor, Computer Science and Robotics in the School of Computer Science, Carnegie Mellon University)

Title: Still Training GANs in 2025?

Abstract:

Recent advances in text-to-image synthesis have led to a shift in the dominant generative modeling paradigms. While GANs were once the de facto choice, models such as DALL·E and Stable Diffusion have popularized autoregressive and diffusion models for large-scale synthesis. This shift raises several questions: Are GANs still relevant? What makes them difficult to train? And what role can they play in the current generative modeling ecosystem? In this talk, I will discuss key challenges in GAN training. I will also present recent advances—both from our group and the broader community—that enable the scaling of GANs for practical applications, including efficient text-to-image synthesis, 4K image upsampling, text-guided image-to-image translation, and the distillation of diffusion models into single-step conditional GANs.

Bio:

Jun-Yan Zhu is the Michael B. Donohue Assistant Professor of Computer Science and Robotics at CMU’s School of Computer Science. Prior to joining CMU, he was a Research Scientist at Adobe Research and a postdoc at MIT CSAIL. He obtained his Ph.D. from UC Berkeley and B.E. from Tsinghua University. He studies computer vision, computer graphics, and computational photography. His current research focuses on generative models for visual storytelling. He is the recipient of the Samsung AI Research of the Year, the Packard Fellowships for Science and Engineering, the NSF CAREER Award, the ACM SIGGRAPH Outstanding Doctoral Dissertation Award, and the UC Berkeley EECS David J. Sakrison Memorial Prize for outstanding doctoral research, among other awards.