DMAD: Distribution Matching as Adversarial Distillation for Fast Visual Generation
DMAD replaces DMD’s auxiliary student-fitted diffusion model with discriminator-based log-density ratio learning.
The paper recasts distribution matching as adversarial classification, using two discriminator heads on a shared backbone to compare real data and teacher samples against the student. Its authors say the resulting linear logit losses recover the DMD distribution-matching gradient at the discriminator optimum. They also add gap-based reweighting to adjust teacher supervision across noise levels. Reported results include one-step ImageNet-64x64 FID of 1.04 and four-step gains on SDXL, Wan2.1-T2V-14B, and MiniMax-H3-33B audio-video generation. ArXiv · AI/CL/LG's note
The paper recasts distribution matching as adversarial classification, using two discriminator heads on a shared backbone to compare real data and teacher samples against the student. Its authors say the resulting linear logit losses recover the DMD distribution-matching gradient at the discriminator optimum. They also add gap-based reweighting to adjust teacher supervision across noise levels. Reported results include one-step ImageNet-64x64 FID of 1.04 and four-step gains on SDXL, Wan2.1-T2V-14B, and MiniMax-H3-33B audio-video generation. ArXiv · AI/CL/LG's note
score 5