A new arXiv paper introduces DMAD (Distribution Matching as Adversarial Distillation), which replaces the auxiliary score-fitting diffusion model required by DMD with two discriminator heads that learn log-density ratios directly. The method achieves state-of-the-art few-step results across image, video, and audio-video generation benchmarks. Code, models, and demos are released publicly.
A new arXiv paper introduces Projected Distribution Matching Distillation (PDMD), which fixes training instability in Distribution Matching Distillation for video diffusion models by filtering out critic errors via a projection. PDMD needs only a one-line code change and improves quality at 4 NFE on Wan2.1 and MiniMax-H3. Code and models are publicly available, making it directly usable by practitioners distilling video models.
A new arXiv paper introduces Projected Distribution Matching Distillation (PDMD), which fixes training instability in Distribution Matching Distillation for video diffusion models by filtering out critic errors via a projection. PDMD needs only a one-line code change and improves quality at 4 NFE on Wan2.1 and MiniMax-H3. Code and models are publicly available, making it directly usable by practitioners distilling video models.
An arXiv paper introduces Elastic Forcing, a post-training method for few-step autoregressive video generation that learns the rollout distribution directly from reference videos, replacing Distribution Matching Distillation's bidirectional diffusion teacher and online fake-score model. It minimizes MMD in frozen self-supervised video representation spaces using a hybrid Nyström–Monte Carlo estimator, improving VBench Total from 83.80 to 84.64 at 17 FPS with a 1.3B model and enabling 14B post-training on eight H200 GPUs.
An arXiv paper introduces Elastic Forcing, a post-training method for few-step autoregressive video generation that learns the rollout distribution directly from reference videos, replacing Distribution Matching Distillation's bidirectional diffusion teacher and online fake-score model. It minimizes MMD in frozen self-supervised video representation spaces using a hybrid Nyström–Monte Carlo estimator, improving VBench Total from 83.80 to 84.64 at 17 FPS with a 1.3B model and enabling 14B post-training on eight H200 GPUs.
Researchers propose DUET, a distillation method that combines trajectory-level and distribution-level distillation experts to achieve high-quality, diverse two-step video generation. The approach assigns an sCM expert to the high-noise step and a DMD expert to the low-noise step, reconciling the quality-diversity trade-off.
Researchers propose DUET, a distillation method that combines trajectory-level and distribution-level distillation experts to achieve high-quality, diverse two-step video generation. The approach assigns an sCM expert to the high-noise step and a DMD expert to the low-noise step, reconciling the quality-diversity trade-off.
An arXiv paper introduces a unified framework for distributional training of one-step visual generators, separating distribution modeling from matching discrepancy via Wasserstein gradient flow. It yields MGFlow, which achieves state-of-the-art ImageNet 256x256 results and converts FLUX.2 [klein] 4B into a one-step text-to-image generator that beats the original four-step model.
An arXiv paper introduces a unified framework for distributional training of one-step visual generators, separating distribution modeling from matching discrepancy via Wasserstein gradient flow. It yields MGFlow, which achieves state-of-the-art ImageNet 256x256 results and converts FLUX.2 [klein] 4B into a one-step text-to-image generator that beats the original four-step model.
An arXiv paper introduces DyMD, a Distribution Matching Distillation framework that adapts teacher supervision and critic fitting to preserve interaction dynamics in few-step video world models. It distills a 14B teacher into a 1.3B four-step student, improving embodied benchmarks and doubling downstream action-planning success versus base DMD.
An arXiv paper introduces DyMD, a Distribution Matching Distillation framework that adapts teacher supervision and critic fitting to preserve interaction dynamics in few-step video world models. It distills a 14B teacher into a 1.3B four-step student, improving embodied benchmarks and doubling downstream action-planning success versus base DMD.