AvatarVerse: High-quality & Stable 3D Avatar Creation from Text and Pose (AAAI2024)

Huichao Zhang, Bowen Chen, Hao Yang, Liao Qu, Xu Wang, Li Chen, Chao Long, Feida Zhu, Daniel K. Du, Shilei Wen

| Project Page | Paper |

Installation

Prepare running enviroment:

• You must have an NVIDIA graphics card with at least 20GB VRAM and have CUDA installed.
• Install Python >= 3.7.
• (Optional, Recommended) Create a virtual environment:

python3 -m virtualenv venv
. venv/bin/activate

# Newer pip versions, e.g. pip-23.x, can be much faster than old versions, e.g. pip-20.x.
# For instance, it caches the wheels of git packages to avoid unnecessarily rebuilding them later.
python3 -m pip install --upgrade pip

• Install PyTorch >= 1.12. We have tested on torch-1.10.0+cu113 and torch2.0.1+cu117, but other versions should also work fine.

• Install kaolin library:

# Replace TORCH_VERSION and CUDA_VERSION with your torch / cuda versions
# e.g. https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-2.0.1_cu117.html

pip install kaolin==0.14.0 -f https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-{TORCH_VERSION}_cu{CUDA_VERSION}.html

• Install torch-scatter library. Change the version according to your torch version, more information can be found here:

pip3 install -v torch-scatter -f https://data.pyg.org/whl/torch-2.0.1+cu117.html

• run the setup_env.sh scripts.

bash setup_env.sh

• optional: install xformers library, the installation of xformers may take some time(sometimes more than 30 minutes):

pip3 install -v git+https://github.com/facebookresearch/xformers.git@main#egg=xformers

Model Card

Our models are provided on the Huggingface Model Page with the OpenRAIL license.

Model	Base Model	Resolution
controlnet-v1.5-densepose	Stable Diffusion 1.5	512x512

Note that you need download the model and config the model path in config.

coarse_model_and_render = dict(
    diffusion='controlnet',  
    diffusion_path = '/your/path/to/control_dense_pose.ckpt',
    ...

Quickstart

set the text prompt in config:

data = dict(
    ...
    text='Elsa in Frozen Disney',  

and run:

python3 main.py --config configs/diffusion/diffusion-3d-controlnet.py --render_test --export_mesh --i_print=100 --i_weights=500 --mixed_precision fp16

GPU Memory

We set different configurations for variable GPU memory, and the best results in our paper requirements large GPU memory(~70GB), you can also use less GPU memory(~20GB) or the default configuration(~32GB) for avatar generation without much degradation in quality.

The configuration of our model is in config. And the gpu memory requirement can be adjusted by changing the coarse_model_and_render, diffusion_config and batch size in the config file:

Different configuration are listed as following, you can choose one from the [Low, Medium, High] list for each parameter, and the GPU memory required are ~20G, ~32G, ~70G respectively.:

coarse_model_and_render = dict(
    ...
    num_voxels=[120/120/160]**3,           # expected number of voxel
    num_voxels_base=[120/120/160]**3,      # to rescale delta distance
    ...
    rgbnet_dim=[6/6/12],                 # feature voxel grid dim
    rgbnet_depth=3,               # depth of the colors MLP 
    rgbnet_width=[64/64/128],             # width of the colors MLP
    )

diffusion_config = dict(
    ...
    controlnet=dict(
        height=[256/384/512],
        width=[256/384/512]
    )
)
coarse_train = dict(
    ...
    N_img=[2/3/4],                   # coarse stage batch size 
  )
fine_train = dict(
    dict(
    ...
    N_img=[2/4/8]

The generated results are shown below with different configuration. Only some local details are lost, such as the geometric details of the hair on the top of the head, or the texture on the dress:

Partial Avatars

With our densepose guidance, we can generate partial avatars with only few parts guidance, now we support full, shoulder and head mode, this can be changed in configuration file config:

coarse_model_and_render = dict(
    ...
    avatar_type='head',  # full shoulder head

Acknowledgement

This repository is heavily based on DirectVoxGO, Controlnet. We would like to thank the authors of these work for publicly releasing their code.

Citation

@article{zhang2023avatarverse,
  title={Avatarverse: High-quality \& stable 3d avatar creation from text and pose},
  author={Zhang, Huichao and Chen, Bowen and Yang, Hao and Qu, Liao and Wang, Xu and Chen, Li and Long, Chao and Zhu, Feida and Du, Kang and Zheng, Min},
  journal={arXiv preprint arXiv:2308.03610},
  year={2023}
}