Wooyeol Lee

I am a first-year MS/Ph.D. student in the Department of Biomedical Sciences at Seoul National University in South Korea. Building upon my undergraduate foundation in computer science, I am continuing my research at the Artificial Intelligence & Biomedical Informatics Lab, advised by Prof. Kyungsu Kim.

My research is driven by the ultimate vision of developing a comprehensive biomolecular world model. Recognizing that the critical bottleneck in this domain is data, my primary focus is on resolving the challenges of data scaling. I focus on the acquisition, processing, and utilization of massive-scale, high-quality structural biology data. To build this robust data foundation, I am actively exploring advanced technologies, including Cryo-EM, to push the boundaries of structurally conditioned protein design and generative AI.

Email  /  CV  /  Github  /  Scholar  /  LinkedIn

profile photo

Research

My research focuses on advancing generative AI, with a core emphasis on macromolecular design and achieving data scaling for biomolecular world models.

ace On the Collapse of Generative Paths: A Criterion and Correction for Diffusion Steering
*, *, , Wooyeol Lee, , , , †,
arXiv preprint, 2025
paper / project page

Ratio-of-densities steering in diffusion models breaks via Marginal Path Collapse. We introduce ACE, a provably stable correction that prevents collapse, cuts error >5x, and enables reliable molecular design.

isac ISAC: Training-Free Instance-to-Semantic Attention Control for Improving Multi-Instance Generation
*, Wooyeol Lee*, *,
arXiv preprint, 2025
paper

Aligning guidance strategy with instance-to-semantic dynamics, where instance layouts form first then semantic details are refined, is crucial for effective multi-instance generation.

trace TRACE: Your Diffusion Model is Secretly an Instance Edge Detector
*, *, Wooyeol Lee, , †,
ICLR 2026 Oral
paper

Diffusion models, capable of generating multiple instances, are secretly powerful annotators for instance segmentation.

Website template from Jon Barron's website. I appreciate his work in creating such a clean and simple design.