Dear all,
You are invited to attend the next talk in the CV-DL Seminar Series.
Omer Dahary
TAU
Sunday, June 7, 2026
11:15-12:30
Room 508, Amir Building
Title:
Toward Meaningful Diversity in Text-to-Image Models
Abstract:
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.
Bio:
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.
We look forward to seeing you there.