Speaker
Nupur Kumari is a Ph.D. student at Carnegie Mellon University, advised by Jun-Yan Zhu. Her research is in generative models, specifically efficient fine-tuning and transfer learning techniques to improve generative models.
Abstract
Advancements in large-scale generative models represent a watershed moment. These models can generate a wide variety of objects, styles, and scenes and their compositions. However, as end users, we often wish to synthesize specific concepts from our own personal lives. Furthermore, these models are trained on a fixed snapshot of available data while new content is created all the time. As a result, these models often lack the capability to generate, with sufficient fidelity, personalized or unique objects that are not within their pre-defined concept space. Nupur will talk about her research on the customization of text-to-image diffusion models to add multiple personal and unique concepts. She will also briefly discuss their recent exploration into ablating (removing) copyright and memorized concepts from the text-to-image diffusion models.