You can deploy this model and start earning money today!

wzhangek/
CoDMD

Wan-AI/Wan2.1-T2V-1.3B

finetune

cc-by-nc-sa-4.0

Views

No views yet

CoDMD: Copula-Aware Distribution Matching Distillation for Video Generation

Paper | Project Page | Code

CoDMD distills Wan2.1 text-to-video diffusion models into 4-step generators while preserving joint dependency structures across video frames and samples through a copula-aware distillation loss.

This repository provides the official generator-only checkpoints for:

Checkpoint	Base model	Sampling steps	File
CoDMD-1.3B	Wan2.1-T2V-1.3B	4	`CoDMD_wan2.1_T2V_1.3B.pt`
CoDMD-14B	Wan2.1-T2V-14B	4	`CoDMD_wan2.1_T2V_14B.pt`

Checkpoint Format

Each model.pt contains:

{"generator": state_dict}

The training-only critic and optimizer state are not included. The released checkpoints are intended for inference with the original CoDMD codebase, whose inference pipeline loads only the generator entry.

Inference

Clone and install CoDMD:

bash

git clone https://github.com/JIA-Lab-research/CoDMD.git
cd CoDMD
pip install -r requirements.txt
pip install easydict
pip install -e .

Download the corresponding Wan2.1 base model, then set model_path in:

configs/wan_dmd_tar.yaml for CoDMD-1.3B
configs/wan_dmd_tar_14b.yaml for CoDMD-14B

Run inference by pointing --checkpoint_folder to the folder containing the downloaded model.pt:

bash

python inference.py \
    --config_path configs/wan_dmd_tar_14b.yaml \
    --checkpoint_folder <HF_REPO_DIR>/CoDMD_wan2.1_T2V_14B \
    --output_folder ./results \
    --prompt_file_path prompts.txt \
    --num_seeds 5

For CoDMD-1.3B, use configs/wan_dmd_tar.yaml and the CoDMD_wan2.1_T2V_1.3B checkpoint folder.

License

The CoDMD release is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. The Wan2.1 base models remain subject to their respective licenses and terms of use.

Citation

bibtex

@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}
}

wzhangek/CoDMD