DMAD: Distribution Matching as Adversarial Distillation for Fast Visual Generation
DMAD distills few-step image and video generators by learning distribution matching with adversarial discriminators.
DMAD turns distribution-matching distillation into classification, training a few-step student from discriminator logit losses without fitting an auxiliary score model. Two heads on a shared backbone separate real and teacher samples from student samples, and gap-based reweighting adapts teacher supervision across noise levels. Reported scores include FID 1.04 for one-step ImageNet-64, FID 14.47 for four-step SDXL on COCO-10K, and VBench 85.15 for four-step Wan2.1-T2V-14B. A four-step MiniMax-H3-33B student wins 79.1% human preference over DMD2 and 84.6% over rCM on joint audio-video, excluding ties.
- Discriminator losses recover the DMD matching gradient without auxiliary score fitting.
- One-step ImageNet-64 generation reaches FID 1.04.
- Four-step Wan2.1-T2V-14B scores 85.15 total on VBench.
- A four-step MiniMax student is preferred over DMD2 and rCM.
Full article212 words · extracted from arxiv.org · click to collapse
Distribution Matching Distillation (DMD) trains a few-step student from the difference between separately estimated target and student scores, so it must keep an auxiliary diffusion model fitted to the student's evolving distribution at extra memory and computation cost. We introduce DMAD, Distribution Matching as Adversarial Distillation, which recasts distribution matching as classification and learns the required log-density ratios directly. Two discriminator heads on a shared backbone distinguish real data and teacher samples from the student's, and linear losses on their logits train the student without auxiliary score fitting. We prove that at the discriminator optimum these losses recover the distribution-matching gradient underlying DMD, through the classical identity linking discriminator logits to log-density ratios. We further introduce gap-based reweighting, which adapts teacher supervision across noise levels from the real-data head's empirical logit gap between real and teacher samples. DMAD reaches a Fréchet Inception Distance (FID) of 1.04 with one-step generation on ImageNet-64x64, 14.47 with four-step SDXL on COCO-10K, and a VBench total score of 85.15 with four-step Wan2.1-T2V-14B, the best values among the compared few-step methods and the multi-step teachers. On MiniMax-H3-33B, our four-step student achieves overall human preference rates of 79.1% over DMD2 and 84.6% over rCM for joint audio-video generation, excluding ties. Our code, models and demos are available at https://yzmblog.github.io/projects/DMAD.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2610.02188