Posterior Matching Distillation improves few-step image generation

Diffusion distillation compresses iterative sampling processes to speed up image generation, but traditional methods often lose sample diversity.

2 min read

Posterior Matching Distillation improves few-step image generation

Diffusion distillation speeds up image and video generation by compressing a pretrained teacher model's slow, iterative sampling process into just a few network evaluations. Traditional distribution-based methods achieve this by forcing the student model's output distribution to match the teacher model. However, this matching process typically relies on reverse-KL objectives, which use a mode-seeking bias — a mathematical property that focuses on dominant modes of the data while ignoring rare variations — leading to a loss of diversity across generated samples.

Key facts

Method
Posterior Matching Distillation (PMD)
Models tested
SDXL and FLUX.2
Model parameters
4B
Sampling steps
4-step methods
arXiv ID
arXiv:2611.26629
Performance Comparison on SDXL and FLUX.2
FID (4-step)
SDXL: best among 4-step methods
FLUX.2 4B: best among 4-step methods
Prompt-alignment and perceptual-quality
SDXL: competitive or leading
FLUX.2 4B: competitive or leading
Research poster detailing Posterior Matching Distillation with method diagrams, architectural training overview, and key benchmark results.
Overview poster illustrating the Posterior Matching Distillation framework, training architecture, and performance results., Source ↗

Introducing Posterior Matching Distillation

Posterior Matching Distillation, known as PMD, is a new approach within the distribution-based distillation family that combines mode-covering posterior sampling with any-step generation. At zero population loss, PMD recovers the teacher's full posterior of clean images — the probability distribution of original images given noisy inputs — at each trained noise level. Its stochastic student supports any-step sampling budgets within the trained timestep grid without requiring additional training.

To learn these posteriors, PMD matches exponentially reweighted averages of student samples conditioned on a fixed noisy state to corresponding teacher posterior means. Under Gaussian corruption, an exponentially tilted posterior mean corresponds to a single teacher evaluation at a shifted input. The research demonstrates that matching tilted means over an open neighborhood uniquely identifies the full posterior.

Results and Performance

PMD supervises the entire conditional distribution through a family of mean constraints, completely removing the need for any auxiliary student-score network. Experiments on SDXL and FLUX.2 4B demonstrate that PMD achieves the best Fréchet Inception Distance, a standard metric used to evaluate the quality of generated images, among four-step methods on both models. It also maintains competitive or leading performance across prompt-alignment and perceptual-quality metrics.

Source: Original announcement

Worth a read?

Report

Comments