DMAD: Distribution Matching as Adversarial Distillation for Fast Visual Generation

DMAD. Distribution Matching as Adversarial Distillation for Fast Visual Generation. By Zhengming Yu and ten others, from Texas A and M University and ByteDance. Published on the first of October. Its code repository is worked out: about two hundred and fifty source files, with a readme and a licence.

Distribution Matching Distillation, or DMD, trains a student that generates in a few steps. Its training signal is the difference between two scores, one for the target and one for the student, each estimated separately. The student's score is the expensive half. Because the student's distribution keeps changing as it trains, DMD has to keep an auxiliary diffusion model fitted to it the whole time, and that model costs extra memory and extra computation.

DMAD, Distribution Matching as Adversarial Distillation, recasts distribution matching as a classification problem. What the student actually needs is a set of log-density ratios, and DMAD learns those ratios directly instead of reaching them through two separately estimated scores. A classifier that can tell distributions apart already carries that ratio in its logits, so the auxiliary score model has nothing left to do.

Concretely, two discriminator heads sit on a shared backbone. One distinguishes real data from the student's samples, the other distinguishes teacher samples from the student's. Linear losses on their logits train the student, with no auxiliary score fitting anywhere. The authors prove that at the discriminator optimum these losses recover the distribution-matching gradient that DMD is built on, through the classical identity linking discriminator logits to log-density ratios.

A second addition is gap-based reweighting. It adapts how much teacher supervision the student receives at each noise level, using the real-data head's empirical logit gap between real samples and teacher samples.

The image results first. With one-step generation on ImageNet at sixty-four by sixty-four, DMAD reaches a Fréchet Inception Distance of one point zero four. With four-step SDXL on COCO-10K the figure is fourteen point four seven. For video, a four-step Wan2.1-T2V-14B student reaches a VBench total score of eighty-five point one five. Each of these is the best value among the compared few-step methods, and among the multi-step teachers too.

Then joint audio-video generation on MiniMax-H3-33B, judged by people. Ties excluded, the four-step student is preferred overall over DMD2 seventy-nine point one percent of the time, and over rCM eighty-four point six percent of the time.

The cost DMAD removes was structural. DMD needed a second model to chase the student because it estimated the student's score on its own. Learning the density ratio by classification, with a proof that the gradient is the same at the optimum, takes that model out of the loop.