Talk Abstract: Generative models are on the cusp of revolutionizing myriad aspects of humanity, having already made a profound impact in the realms of photography, art, and creativity. Among these models, text-to-image diffusion models stand out for their unprecedented capability to translate semantic intent into visually compelling outputs. Characterized by their iterative nature, these models constitute a framework for diverse interventions during training and inference that enables users to manipulate and customize their outputs in unique ways. In this talk, we will dive into the remarkable properties of pre-trained diffusion models—models that have been refined through training on billions of examples and explore how to effectively intervene in the generation process, guiding and tailoring them to unlock limitless possibilities. In particular, I will introduce the concept of personalized generative models, illustrating how models can be adapted to specific user needs with a few examples and address challenges previously considered insurmountable.

Speaker Biography: Kfir Aberman is a Research Scientist leading the Generative AI research effort at Snap Research. His research focuses on generative models for image synthesis, editing, as well as human motion modeling and manipulation. Before joining Snap, Kfir was a Senior Research Scientist at Google. He received his Ph.D. from Tel-Aviv University, advised by Prof. Daniel Cohen-Or. Besides his core responsibilities, Kfir is an active contributor to the graphics and vision community, serving as a committee member and area chair at various industry events and conferences.