CoDMD: Copula-aware Distribution Matching
Distillation for Fast Video Generation

Wenhu Zhang1, Kun Cheng2*, Changyuan Wang3, Shiyao Li5, Yuechen Zhang4, Wenbo Li4, Jiajun Zha1, Jingyi Zhang2, Kang Zhao2, Jiaya Jia1†

1HKUST    2Wan Team, Alibaba Group    3Tsinghua University    4CUHK    5Xidian University
*Project Leader    †Corresponding Author

📄 Paper 💻 Code 🤗 Models

A lightweight relational regularizer that makes your DMD motion-smoother, color-faithful, and instruction-precise.

Student [CoDMD(Ours), DMD, rCM]: 4-step generation, w/o CFG  |  Teacher: Wan-2.1 T2V 14B, 50 steps + CFG (2× compute)

Method

Standard DMD minimizes reverse-KL with coordinate-wise score differences, imposing no constraint on relational geometry across batch elements or temporal frames. CoDMD introduces a copula-aware regularizer that reuses score estimates from the frozen teacher and online fake model to construct pairwise relation matrices, matched through a supplementary distributional objective — requiring no additional networks, datasets, or sampling trajectories.

CoDMD Method Overview

Generated Videos

🎨 More Vivid Colors

🧩 Precise Instruction Following

🏃 Coherent Action Dynamics

🎥 Smooth Camera Motion

VBench Results

CoDMD distills Wan-2.1-T2V models at 1.3B & 14B scales into 4-step students, achieving VBench total scores of 84.46 & 84.87 — outperforming trajectory-based (rCM: 82.81 & 84.05) and distribution-based (DMD: 83.38 & 83.81) methods.

Wan-2.1-T2V 1.3B

VBench Radar - 1.3B Model

Wan-2.1-T2V 14B

VBench Radar - 14B Model

Contact Us

Feel free to contact Wenhu Zhang at wenhuzhang98@gmail.com for any question, cooperation, and communication.

Citation

If you find this work useful, please consider citing:

@misc{zhang2026codmdcopulaawaredistributionmatching,
      title={CoDMD: Copula-aware Distribution Matching Distillation for Fast Video Generation}, 
      author={Wenhu Zhang and Kun Cheng and Changyuan Wang and Shiyao Li and Yuechen Zhang and Wenbo Li and Jiajun Zha and Jingyi Zhang and Kang Zhao and Jiaya Jia},
      year={2026},
      eprint={2606.21982},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.21982}, 
}