This is the code repository of Dynamic Motion Blending for Versatile Motion Editing at CVPR 2025.

📝 arXiv | 🌐 Project Page | 🤗 Hugging Face Space

To run the application, you need to have the following installed:

• Python 3.10
• Required Python packages (specified in requirements.txt)
• git-lfs

1. Clone the Repository:

   git clone https://github.com/emptybulebox1/motionReFit.git
   cd motionRefit

2. Download Checkpoints and SMPL-X Models:

   • You can find them in https://huggingface.co/Yzy00518/motionReFit, where contains all checkpoints and data required for demo.
   • You can also run app.py, which can automatically download all weights from huggingface repo and put them to correct position.

3. Install Python Packages:

   pip install -r requirements.txt

• Before running the application, use git-lfs to pull data from this repo

  git-lfs install
  git-lfs pull

• Then start the Gradio application:

  python app.py

• Comming soon!

STANCE (Style Transfer, Fine-Grained Adjustment, and Body Part Replacement) is a comprehensive motion editing benchmark that encompasses three common types of motion editing tasks.

This task focuses on precise modifications of specific body part movements. For example, “replace a waving right hand with a swinging motion” or “change a walking leg movement to running”. To support this task, we have annotated 13,000 motion sequences. Each sequence is tagged with precise body part masks (UPPER_BODY, LOWER_BODY, BOTH_ARMS, LEFT_ARM, RIGHT_ARM) along with corresponding motion descriptions.

This task is designed to change the stylistic expression of a motion while preserving its semantic content. For example, “transform a calm gesture into an angry one”. We collaborated with experienced motion capture actors to perform multiple emotional and stylistic variations of the same motion. In total, we collected approximately 2 hours of high-quality footage, covering styles such as sexy, angry, and old.

This task involves subtle modifications of motion characteristics such as amplitude and force — for example, “walk faster” or “increase the swing of a hand wave”. To support this, we have constructed 4,500 motion pairs along with editing descriptions. These triples were partially generated using large language models and later validated by experts to ensure quality.

Please download the STANCE dataset from Google Drive. The content inside the download link will be continuously updated to ensure you have access to the most recent data.

The file structure should be like:

dataset/
├── base_motion
│   ├── 000000.pkl
│   ├── 000002.pkl
│   ├── ...
│   └── 029231.pkl
├── regen
│   ├── mask_all.json
│   └── part_annotations.json
├── style_transfer
│   ├── 000009_depressed.pkl
│   ├── 000009_proud.pkl
│   ├── ...
│   └── 000642_sexy.pkl
├── adjustment
│   ├── paired_data_seed0_15_batch0_id3.pkl
│   ├── paired_data_seed0_15_batch0_id5.pkl
│   ├── ...
│   └── paired_data_seed2_13_batch4_id129.pkl
├── split
│   ├── base_motion
│   │   ├── val.txt
│   │   └── test.txt
│   ├── style_transfer
│   │   ├── val.json
│   │   └── test.json  
│   └── adjustment
│       ├── val.json
│       └── test.json  
└── README.md

Explanation of the files and folders of the STANCE dataset:

• base_motion (folder): SMPL-X format motion data from HumanML3D (24746 in total, 20 FPS).
• regen (folder): Motion data for body part replacement is in base_motion.
  • mask_all.json: Annotated body parts, where each key is a HumanML3D ID.
  • part_annotations.json: Text annotations of annotated body parts, where each key is a HumanML3D ID.
• style_transfer (folder): SMPL-X format motion data for motion style transfer (749 in total, 20 FPS).
• adjustment (folder): SMPL-X format motion data for fine-grained motion adjustment (4411 in total, 20 FPS).
• split (folder): Dataset split of different subsets.
  • base_motion (folder): Official Val & Test splits from HumanML3D.
  • style_transfer (folder): Val & Test pairs used in MotionReFit.
  • adjustment (folder): Val & Test pairs used in MotionReFit.
• Data format for .pkl files in base_motion and style_transfer:

  {   
      "body_pose": numpy.ndarray (N, 63),
      "global_orient": numpy.ndarray (N, 3),
      "transl": numpy.ndarray (N, 3),
  }
  

• Data format for .pkl files in adjustment:

  {
      "src": {   
          "body_pose": numpy.ndarray (N, 63),
          "global_orient": numpy.ndarray (N, 3),
          "transl": numpy.ndarray (N, 3),
      },
      "tgt": {   
          "body_pose": numpy.ndarray (N, 63),
          "global_orient": numpy.ndarray (N, 3),
          "transl": numpy.ndarray (N, 3),
      },
      "body_part": body part to edit,
      "text": editing instruction,
  }
  

Note: During the training process of motion style transfer and fine-grained motion adjustment, only the blended motions (generated as BLD(base_motion, adjustment_motion)) are available.

@article{jiang2025dynamic,
  title={Dynamic Motion Blending for Versatile Motion Editing},
  author={Jiang, Nan and Li, Hongjie and Yuan, Ziye and He, Zimo and Chen, Yixin and Liu, Tengyu and Zhu, Yixin and Huang, Siyuan},
  journal={arXiv preprint arXiv:2503.20724},
  year={2025}
}

We adapted some code from other repos in data processing, training, evaluation, etc. Please check these useful repos.

https://github.com/jnnan/trumans_utils
https://github.com/mileret/lingo-release
https://github.com/atnikos/motionfix
https://github.com/GuyTevet/motion-diffusion-model
https://github.com/Mathux/TMR