Yu Liu

DiffusionOPD: A Unified Perspective of On-Policy Distillation in Diffusion Models

Theoretically, the OPD framework is lifted from discrete tokens to continuous-state Markov processes, deriving a closed-form per-step KL objective that unifies both stochastic SDE and deterministic ODE refinement via mean-matching, which provides lower variance and better generality compared to conventional PPO-style p...