We explore how generative models such as VAEs, GANs, and Diffusion models can create images and videos. Beyond generation itself, we develop Editing and Composition techniques that allow users to modify visual content as they intend. Building on these foundations, our work is now moving toward Action Generation and World Models for real-world AI. 

Pretrained AI models often struggle when they are applied to new domains. We study Domain Adaptation and Domain Generalization so that these models can be used effectively across diverse domains. We also develop learning techniques based on representation learning to improve classification and segmentation.

Advances in visual intelligence become most meaningful when they solve real problems. We apply our research to practical domains through funded projects and interdisciplinary collaborations, focusing on Document AI, Fashion AI, and AI for Science.

Document AI

Documents captured in everyday life are often warped, folded, or degraded. We develop Document AI models for automated document processing, including de-warping and restoration techniques for real-world document images. Leveraging the capabilities of large language models, we are also extending this work toward Visual LLMs that understand visual information in documents.

Fashion AI

Clothing carries rich visual information such as fabric texture, material properties, and garment shape. Through government and industry funded projects, we develop AI technologies that understand and generate fabrics and garments, helping to bring digital workflows into the fashion industry.

AI for Science

Many scientific fields now face data that is too complex to analyze by hand. In collaboration with faculty members from the College of Pharmacy, we explore how AI can accelerate scientific discovery, including applications in drug discovery.