Bohyung Han

D2PO: Optimizing Diffusion Samplers via Dynamic Preference

Jinkyu Kim, Jinyoung Choi, Bohyung Han

We propose D2PO (Dynamic Direct Preference Optimization), a principled framework for optimizing diffusion sampling policies with respect to timestep schedules and classifier-free guidance (CFG) weights. Our work is motivated by a fundamental limitation of existing student-teacher…Read more