S307, New Bund Campus, NYU Shanghai
Tuesday, October 07, 2025 - 11:45 - 12:45
Join our PhD and Undergraduate Students Research Spotlight Series. This series provides a platform for students to share their work and connect with the academic community!
About the Speaker: Linxi Xie (Cissy) is a senior student majoring in Computer Science at NYU Shanghai. Her research interests focus on controllability and personalization of generative models.
Abstract: Current AI models for image generation are powerful, but current systems struggle to capture the fine-grained preferences of individual users. A key barrier is the lack of large-scale, real-world datasets that reflect how people actually interact with these models. To address this, we introduce ImageGem, a dataset built from the interactions of 57K users, including 242K customized LoRA models, 3M text prompts, and 5M generated images. This unique combination of prompts, outputs, and model customizations provides, for the first time, a foundation for systematically studying personalization in image generative models such as Stable Diffusion.
In this talk, we will delve deeper into a framework that leverages ImageGem to align diffusion models with individual preferences. Our approach directly edits models in latent weight space, enabling fine-grained personalization without retraining from scratch. We demonstrate that this framework makes generative models more adaptive and reliable across diverse users.