NVIDIA (Hidir Yesiltepe, Julius Berner, Arash Vahdat, Pinar Yanardag) · Diffusion distillation

Posterior Matching Distillation

Posterior Matching Distillation (PMD) is a distribution-based method for compressing pretrained diffusion teachers into students that generate images or video in a few network evaluations. It matches reweighted averages of student samples to teacher posterior means, aiming to avoid the diversity loss of reverse-KL objectives, and needs no auxiliary student-score network. On SDXL and FLUX.2 [klein] 4B, the authors report the best FID among 4-step methods on both models, with competitive or leading prompt-alignment and perceptual-quality scores. No licence or code availability is stated on the page.

Open the source ↗

Similar Research & Safety releases

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.