Distills a diffusion teacher’s conditional image distribution by matching exponentially tilted student means to teacher posterior means. It supports flexible sampling budgets on the trained timestep grid while preserving generation diversity.