Posterior Matching Distillation preserves teacher diversity in few-step generation

ArashVahdat · x · 2026-11-21

A team including Arash Vahdat introduces Posterior Matching Distillation (PMD), a new distribution-based distillation method for few-step generative models. PMD trains a stochastic student whose samples reproduce the teacher's exponentially tilted means, each obtained from a single frozen-teacher evaluation at an input shifted along the tilt direction. They prove that matching these tilted means over a neighborhood uniquely determines the teacher's full posterior over clean images, yielding exceptional sample diversity across seeds for the same prompt.

Related event: NVIDIA Proposes Posterior Matching Distillation for Few-Step Generation(2 posts)→

Original post →

More from Multimodal

Multimodal channel →