Semantically Aligned Multi-person Image Personalization

The School of EECS is hosting the following MPhil Progress Review 1 seminar: 

Semantically Aligned Multi-person Image Personalization

Speaker: Longpeng Xu
Host / Chair: Dr. Yadan Luo

Abstract:

Current text-to-image models struggle with multi-person image personalization, often failing at semantic alignment—correctly matching detailed texts to the right individuals and attend to overall semantics of the texts. To address this, we introduce a novel framework that generates high-fidelity, multi-person images. Our method achieves this through two-way alignment between a powerful Multimodal Large Language Model for semantic understanding and a Diffusion Transformer for high-fidelity, identity-preserving generation.

Bio: Longpeng Xu is an MPhil student at the School of EECS at University of Queensland, under the supervision of Dr. Priyanka Singh and Prof. Xue Li. He obtained Master of Data Science and Bachelor of Commerce at University of Queensland and University of Western Australia, respectively. His research interests include multimodal understanding and generation.