Yang, S. (CSE) – Beyond Image Editing: Building Generalized Image Customization Systems

Current generative vision models struggle with image customization that requires multi-step reasoning or real-world knowledge. This proposal introduces generalized image customization, enabling systems to execute complex, inferential modifications rather than just simple edits. The research focuses on the foundational framework required for this generalization, specifically high-quality training data, scalable evaluation benchmarks, self-improving training paradigms that reduce reliance on paired annotations, and unified multi-modal architectures. Building on two completed studies in data quality and evaluation, this proposal outlines two future research directions to develop capable, annotation-efficient, and reasoning-native image customization systems.

Event Host: Siwei Yang, Ph.D. Student, Computer Science and Engineering

Advisor: Cihang Xie

Zoom- https://ucsc.zoom.us/j/3852138080?pwd=Z0MyTVM2WjdCbEM4OXVxWUhhei84dz09