CHARM

📃 Paper • 💻 [Github Repo] • 🌐 Project Page

This repository contains the test dataset presented in the paper CHARM: Control-point-based 3D Anime Hairstyle Auto-Regressive Modeling.

CHARM is a novel parametric representation and generative framework for anime hairstyle modeling.

Usage

You can download the files directly from this repository or use the Hugging Face datasets library:

from huggingface_hub import hf_hub_download, list_repo_files

# Get list of all files in repo
files = list_repo_files(repo_id="hyz317/CHARM", repo_type="dataset")

# Download each file
for file in files:
    file_path = hf_hub_download(
        repo_id="hyz317/CHARM",
        filename=file,
        repo_type="dataset",
        local_dir='./test_cases'
    )

AnimeHair Dataset

Due to policy restrictions, we are unable to redistribute the raw 3D models of the training dataset. However, you can download the VRoid dataset by following the instructions provided in PAniC-3D. Please note: When downloading the VRoid dataset, replace the metadata.json file mentioned in the PAniC-3D instructions with the file from this link. All other steps should follow the original PAniC-3D guide.

In place of the raw data, we are providing the preprocessing scripts and training data list.

First, install Blender and download the VRM Blender Add-on. Then, install the add-on using the following command:

blender --background --python blender/install_addon.py -- VRM_Addon_for_Blender-release.zip

Next, execute the Blender script to separate the hair meshes from the 3D models:

cd blender
python distributed_uniform.py --input_dir /PATH/TO/YOUR/VROIDDATA --save_dir /PATH/TO/YOUR/SAVEDIR --workers 32

• /PATH/TO/YOUR/VROIDDATA: Specify the directory where you downloaded the VRoid dataset.
• /PATH/TO/YOUR/SAVEDIR: Specify the directory where the separated hair models will be saved.

After separating the hair, run the following script to post-process the hairstyles and convert them into a template format:

python process_hair.py --input_dir /PATH/TO/YOUR/SAVEDIR

• /PATH/TO/YOUR/SAVEDIR: This should be the same directory where the separated hair models were saved in the previous step.

Finally, sample point clouds from the processed hair models:

python sample_hair.py --input_dir /PATH/TO/YOUR/SAVEDIR

• /PATH/TO/YOUR/SAVEDIR: Again, use the directory containing the post-processed hair models.

Citation

If you find our work useful, please kindly cite:

@article{he2025charm,
  title={CHARM: Control-point-based 3D Anime Hairstyle Auto-Regressive Modeling},
  author={He, Yuze and Zhou, Yanning and Zhao, Wang and Ye, Jingwen and Bai, Yushi and Xiao, Kaiwen and Liu, Yong-Jin and Sun, Zhongqian and Yang, Wei},
  journal={arXiv preprint arXiv:2509.21114},
  year={2025}
}