Diffusion models often need many network evaluations to turn noise into an image. Distillation seeks to compress that iterative process into a smaller number of steps, but speed alone does not guarantee that the student retains the range of outputs represented by its teacher. Posterior Matching Distillation (PMD) is designed to address both concerns: fewer evaluations and a better match to the teacher’s conditional distribution.
Matching more than a dominant mode
The paper argues that distribution-distillation objectives based on reverse KL can be mode-seeking. In practice, that bias may encourage a student to focus on the teacher’s dominant modes while representing less frequent alternatives poorly. PMD instead aims to learn the teacher’s posterior distribution over clean images for a given noisy state.
For each fixed noisy input, the method draws student samples and forms exponentially reweighted averages. It then matches these averages to corresponding posterior means from the teacher. Under Gaussian corruption, an exponentially tilted posterior mean can be obtained with one teacher evaluation at a shifted input. The authors also prove that matching tilted means across an open neighborhood uniquely identifies the full posterior. In this way, a family of mean constraints can supervise a conditional distribution without an auxiliary student-score network.
What the method is intended to offer
- Posterior recovery: At zero population loss, PMD recovers the teacher’s full clean-image posterior at each trained noise level.
- Flexible step budgets: Its stochastic student can sample with different budgets within the trained timestep grid, without additional training for each budget.
- Reported evaluation: The abstract describes experiments on SDXL and FLUX.2 [klein] 4B and claims the best FID among four-step methods on both models.
Why it matters—and what remains unclear
The proposal is notable because it treats distillation as a problem of recovering a conditional distribution, rather than merely matching a single prediction. If its results generalize, that framing could help few-step generation retain more variation while giving users flexibility to trade compute for sampling steps. The supplied abstract, however, is truncated during its final experimental claim, so it does not provide the complete comparison or supporting details.
There is also a specific result that warrants checking. A comment on the Hugging Face paper page says a FLUX.2 COCO-10k FID in one table is 17.47, while another table reports 24.12 for PMD; the commenter notes that diversity scores appear to match across the entries. This is a reported discrepancy, not an independently verified correction. Until the paper tables or an author response clarify it, the exact FLUX.2 FID comparison should be treated cautiously rather than repeated as settled evidence.
Source: Hugging Face Daily Papers
Comments
Checking sign-in status...
Loading comments...