Modern text-to-image models achieve strong visual fidelity and prompt alignment, but often at the cost of generative diversity. In this talk, I will present two complementary approaches to this problem: a simple inference-time method that achieves rich diversity by intervening in the model’s internal representations, and a new task formulation of controlled diversity, where users explore structured image galleries through meaningful semantic variations.
Omer Dahary is a PhD student in Computer Science at Tel Aviv University, advised by Daniel Cohen-Or. His research focuses on generative models, with an emphasis on improving their ability to align with user control. He received dual B.Sc. degrees in Computer Science and Mathematics from the Technion through the Rothschild Excellence Program, and subsequently earned his M.Sc. at the Technion under the supervision of Alex Bronstein.