CS Events

PhD Defense

Deep Generative Models: Controllability, Efficiency, and Beyond


Download as iCal file

Monday, July 01, 2024, 11:00am - 01:00pm


Speaker: Ligong Han

Location : CoRE 301


Prof. Dimitris Metaxas (Chair)

Prof. Vladimir Pavlovic

Prof. Hao Wang

Prof. Qiang Liu (external)

Event Type: PhD Defense

Abstract: The rapid evolution of deep generative models has unlocked unprecedented capabilities in AI, pushing the boundaries of creativity, personalization, and efficiency. This dissertation explores two works that exemplify the integration of controllability and efficiency in the domain of generative AI, specifically focusing on text-to-image diffusion models. First, I discuss our parameter-efficient approach for personalizing text-to-image diffusion models. By optimizing the singular values of weight matrices, this method facilitates the adaptation of pre-trained models to new tasks and capabilities with minimal data, enabling concept composition and image editing. The second work focuses on leveraging diffusion inversion techniques for controlled image editing and solving inverse problems. By developing approximate algorithms for optimization-based methods, we achieve not only time efficiency but also enhanced performance.

Contact  Professor Dimitris Metaxas

Join Zoom Meeting
Join by SIP

Meeting ID: 239 465 5136
Password: 640958
One tap mobile
+16465588656,,2394655136# US (New York)
+16469313860,,2394655136# US