Causal Motion Diffusion Models (CMDM)

This repository contains materials developed by LY Corporation and is temporarily open-sourced for the purpose of our reseach project.

• Temporary Release: This repository is temporarily available as open-source. Therefore this repository may be turn into read-only or private anytime.
• Attribution: All code and materials in this repository are owned by LY Corporation.

Project Overview

Code of the paper "Causal Motion Diffusion Models for Autoregressive Motion Generation" (CVPR 2026).

⚙️ Getting Started

Installation, pre-trained models, and data

1. Python Environment using uv

uv sync

2. Models and Dependencies

Download Evaluation Models

1. Download glove to glove folder and t2m_evaluators to checkpoints folder from MARDM repository for robust 67-dim evaluation.

2. Download t2m_evaluators from TM2T repository to checkpoints and rename its text_mot_match folder to text_mot_match_full for 259-dim evaluation.

Download Pre-trained Models

Download pre-trained models from huggingface and put them in checkpoints/t2m/.

3. Obtain Data

Download HumanML3D Dataset from the repository of HumanML3D. Unzip and locate it in the datasets folder.

The whole directory should be look like this:

CMDM
│   README.md
│   pyproject.toml
|   ...
|
└───glove
└───train
└───...
│   
└───datasets
|    └───HumanML3D
|       └───new_joint_vecs
|       └───...
│   
└───checkpoints
     └───t2m
          └───pretrained_tmr
          └───pretrained_vae
          └───pretrained_dit
          └───text_not_match
          └───text_not_match_clip
          └───text_not_match_full

🎬 Demo

Demo scripts

Text-to-motion generation (default)

uv run python -m sample.demo \
  prompt_csv=sample/sample_prompt.csv \
  exp=my_demo \
  autoregressive=false \
  num_samples=5

When autoregressive=false and prompt_csv is provided, the script generates num_samples results for each (prompt, length) row in the CSV.

Outputs are saved in generations/<exp>/ as:

• sample_XX_repXX.mp4
• sample_XX_repXX.npy

Here, the first XX is the prompt-length pair index, and the second XX is the repetition index.

Autoregressive long-horizon generation

uv run python -m sample.demo \
  prompt_csv=sample/sample_prompt.csv \
  exp=my_demo_ar \
  autoregressive=true \
  num_samples=5

When autoregressive=true and prompt_csv is provided, each CSV row is used as the next segment in one long sequence, generated in order with causal conditioning from prior segments. num_samples controls how many full long sequences are sampled.

Outputs are saved in generations/<exp>/ as:

• long_sample_repXX.mp4
• long_sample_repXX.npy

📊 Evaluate Pre-trained CMDM models

Evaluate pre-trained models

Evaluate MAC-VAE

uv run python -m eval.eval_ae name=pretrained_vae

Evaluate Causal-DiT

Robust evaluation Without redundant features.

uv run python -m eval.eval_ae_dit name=pretrained_dit ae_name=pretrained_vae feature_dim=67

Use feature_dim=259 for general 259-dimensional evaluation.

📖 Train CMDM

Train CMDM models

Train/eval entrypoints are Hydra-based. Default configs are in conf/, and CLI overrides use key=value (for example, gpu=0 name=MAC_VAE).

Train Part-TMR

uv run python -m Part_TMR.scripts.train

Train MAC-VAE

uv run python -m train.train_ae name=MAC_VAE

Train Causal-DiT

uv run python -m train.train_ae_dit name=Causal_DiT

🎆 Evaluate CMDM models

Evaluate CMDM models

Evaluate MAC-VAE

uv run python -m eval.eval_ae name=MAC_VAE

Evaluate Causal-DiT

uv run python -m eval.eval_ae_dit name=Causal_DiT

Acknowledgement

Some parts of our code are based on ACMDM and other third-party software listed in NOTICE.txt.

Citation

@inproceedings{yu2026causal,
      title={Causal Motion Diffusion Models for Autoregressive Motion Generation},
      author={Yu, Qing and Watanabe, Akihisa and Fujiwara, Kent},
      booktitle={CVPR},
      year={2026}
  }

Contributions

As this project is temporarily open-sourced, we are not accepting contributions. For feedback or inquiries, please open an issue in this repository.

License

This code is dedicated to the public domain under CC0 1.0. You may copy, modify, and distribute it without restriction, and the authors make no warranties or guarantees regarding its use.

Additionally, this repository contains third-party software. Refer NOTICE.txt for more details and follow the terms and conditions of their use.