Summary

PDMD, or Projected Distribution Matching Distillation, is a method and implementation for faster diffusion generators. A projection removes the component of the training direction associated with estimated critic error, without adding another network or model pass.

Key Features

Distribution-matching distillation.
Critic-error projection.
One-line target modification.
No extra network.
No additional model pass.
No auxiliary loss.
Few-step generator training.
Public code and weights.

About PDMD


Researchers can incorporate a lightweight change into existing distillation pipelines and evaluate the supplied model weights. Core capabilities include Distribution-matching distillation, Critic-error projection, One-line target modification, and No extra network.


Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include No additional model pass, No auxiliary loss, Few-step generator training, and Public code and weights.

Get more likes & reach the top of search results by adding this button on your site!

Embed button preview - Light theme
Embed button preview - Dark theme
TurboType Banner

Subscribe to the AI Search Newsletter

Get top updates in AI to your inbox every weekend. It's free!