Organizations: MoE Key Lab of Artificial Intelligence, X-LANCE Lab, Shanghai Jiao Tong University · State Key Laboratory of General Artificial Intelligence, BIGAI · Shanghai Innovation Institute · Tencent Hunyuan · Nanyang Technological University
Speech foundation models built on flow matching handle generation, editing, and processing in one system, but every request costs 20-50 iterative steps, and classifier-free guidance (CFG) doubles each of them. Distribution matching distillation (DMD) trains a student that approaches the teacher's quality in a few steps, yet existing speech distillation work applies it to generation alone. We show that a distribution-matching objective does not serve every task: in separation, the noised states of erroneous student outputs deviate far from the teacher's training distribution, where its prediction is unreliable. Route-DMD therefore makes the objective task-dependent, routing the affected tasks to a regression loss and keeping decoupled distribution matching elsewhere. The student is initialized by consistency distillation, and teacher rollouts use adaptive projected guidance (APG) instead of CFG. Without CFG, the 4-step student matches the 32-step teacher on generation, editing, and processing tasks at a 4.5× end-to-end speedup.
Figures & tables
Figure 1: Overview of Route-DMD. Stage 1 uses consistency distillation for initialization. Stage 2 routes separation samples to a regression loss and other samples to Decoupled DMD.
Table 1: Comparison with prior systems and the teacher on representative tasks. The best results are in bold , and the second-best results are underlined .
Figure 2: Ablation of the task-routed mechanism on Libri2Mix. Separation metrics vs. Decoupled DMD updates for the regression loss (w/ routing), Decoupled DMD with the teacher queried on a noised ground-truth target rather than a noised student output (w/ noised GT), and plain Decoupled DMD (w/o routing).
System
N-MOS ↑
I-MOS ↑
RTF ↓
AuK (32 steps)
4.23 ± 0.10
4.50 ± 0.09
0.4333
w/ Route-DMD (4 steps)
4.34 ± 0.10
4.57 ± 0.09
0.0958
Table 2: Results of the subjective evaluation are reported as mean ± 95% confidence interval (CI) over all ratings. End-to-end RTF is measured on a single NVIDIA H20 GPU.
Figure 3: Ablation of initialization on SeedTTS-Eval. Each curve corresponds to a student initialized before DMD by consistency distillation (CD), ODE regression (ODE), meanflow distillation, or no initialization.
Figure 4: Ablation of guidance in the DMD teacher rollouts. SIM and WER on SeedTTS-Eval versus DMD updates.
Flow and diffusion models suffer from slow inference due to computationally expensive numerical integration. Distillation provides a promising way for a student model to learn from a teacher's dynamics, enabling one-step or few-step generation. However, existing methods often depend on curated distillation datasets, costly teacher rollouts, or auxiliary proxy networks, which complicate model training and scaling. In this work, we propose Consistent Distribution Matching, a simulation-free and data-free distillation method for accelerating diffusion and flow models while preserving strong generative capacity. Our key insight is to unify sample generation and score estimation with one student network. Thus, our framework uses only two models, a frozen teacher and a trainable student, and optimizes one objective. We prove that minimizing our objective indicates Wasserstein convergence of the student flow-map pushforwards to the teacher marginals. On ImageNet 256×256, our method attains an FID of 2.04 with a single function evaluation (1-NFE) and a 4-NFE FID of 1.37 within 40 epochs of training, surpassing the state-of-the-art distillation baselines without data. Our code code and model are available at https://consistentdmd.github.io/.
Yuxiang Fu, Qi Yan, Zike Wu +4
University of British Columbia · Vector Institute for AI · Canada CIFAR AI Chair
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.
Zhengming Yu, Junkun Yuan, Haotian Yang +8
Texas A&M University · Intelligent Creation, ByteDance
Distribution Matching Distillation (DMD) distills multi-step diffusion models into few-step generators by minimizing reverse KL divergence. Despite its strong performance, DMD guides samples by local density alone, which introduces two limitations: teacher scores can be unreliable in regions of low teacher density, and the mode-seeking behavior of reverse KL can collapse the student onto a subset of the teacher modes. We first connect DMD to Wasserstein gradient flow, showing that its gradient follows from a fixed-point formulation with integral KL (IKL) as the energy functional. This perspective motivates augmenting IKL with an optimal transport term that addresses these limitations by guiding outlier samples toward the target and promoting mode coverage. We realize this term using Sinkhorn divergence estimated from finite batches. In practice, however, errors in the estimated Sinkhorn velocity can perturb in-domain samples for which IKL guidance is already accurate. To this end, we introduce sDMD, which adaptively activates the Sinkhorn correction based on discriminator outputs, focusing on outlier samples where IKL may be unreliable. With 2--4 sampling steps on FLUX.1-dev, Qwen-Image-20B, and Wan2.1-14B, sDMD outperforms distillation baselines on GenEval, DPG, WISE, and VBench while improving sample diversity over DMD2.