Xuangeng Chu¹  Tatsuya Harada^1,2
1The University of Tokyo, ²RIKEN AIP

🤩 NeurIPS 2024 🤩

GAGAvatar reconstructs controllable 3D head avatars from single images.
GAGAvatar achieves one-shot 3DGS-based head reconstruction and ⚡️real-time⚡️ reenactment.
🔥 More results can be found in our Project Page. 🔥

Installation

Clone the project

git clone --recurse-submodules git@github.com:xg-chu/GAGAvatar.git
cd GAGAvatar

Build environment

conda env create -f environment.yml
conda activate GAGAvatar

Install the 3DGS renderer

What’s the difference between this version and the original 3DGS?

• We changed the number of channels so that 3D Gaussians carry 32-dim features.
• We changed the package name to avoid conflict with the original Gaussian splatting.

git clone --recurse-submodules git@github.com:xg-chu/diff-gaussian-rasterization.git
pip install ./diff-gaussian-rasterization
rm -rf ./diff-gaussian-rasterization

Prepare resources

Prepare resources with:

bash ./build_resources.sh

Also prepare resources for GAGAvatar_track using:

cd core/libs/GAGAvatar_track
bash ./build_resources.sh

Quick Start Guide

Driven by another image:

# This will track the images online, which is slow.
python inference.py -d ./demos/examples/2.jpg -i ./demos/examples/1.jpg

Driven by a tracked video:

python inference.py -d ./demos/drivers/obama -i ./demos/examples/1.jpg

Driven by a tracked image_lmdb:

python inference.py -d ./demos/drivers/vfhq_demo -i ./demos/examples/1.jpg

To test the inference speed, refer to the speed_test() function in inference.py.

To test your own images online, refer to lines 52-55 in inference.py.

To test your own driving sequences (videos/images), refer to GAGAvatar_track and demo sequences to build your own driving sequence.

Training Guide

You can use the pre-trained model directly, but if you need to retrain on your data:

Step 1: Building the image LMDB

Build img_lmdb yourself.

All the images should be cropped as inference. (Refer to line 218 in core/libs/GAGAvatar_track/engines/engine_core.py)

Dump images using core/libs/utils_lmdb.py, there is also an API for building lmdb: dump(key_name, payload), payload should be tensor with (3, 512, 512), in [0, 255].

015252 is video id (used when sampling), 99 is frame id (0 is the first frame, other frames id can be discontinuous).

img_lmdb:
    '015252_99' : image payload

Step 2: Track the image LMDB

Using track_lmdb.py in GAGAvatar_track, you should get a optim.pkl.

optim.pkl:
- dict_keys(['000000_0', …])
    - "000000_0": dict_keys(['bbox', 'shapecode', 'expcode', 'posecode', 'eyecode', 'transform_matrix'])

Step 3: Split the dataset

Build dataset.json yourself, it should contain the keys in img_lmdb and optim.pkl.

dataset.json: {
    "train": ["000000_0", "000000_5", ..., '001384_654'],
    "val":   ["015209_0", ..., "015218_7"], 
    "test":  ["015203_0", ..., "015252_139"]
}

Step 4: Modify the config and train

python train.py --config gaga --dataset vfhq

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{
    chu2024gagavatar,
    title={Generalizable and Animatable Gaussian Head Avatar},
    author={Xuangeng Chu and Tatsuya Harada},
    booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
    year={2024},
    url={https://openreview.net/forum?id=gVM2AZ5xA6}
}

Acknowledgements

Some part of our work is built based on FLAME, StyleMatte, EMICA and VGGHead. The GAGAvatar Logo is designed by Caihong Ning. We also thank the following projects for sharing their great work.

• GPAvatar: https://github.com/xg-chu/GPAvatar
• FLAME: https://flame.is.tue.mpg.de
• StyleMatte: https://github.com/chroneus/stylematte
• EMICA: https://github.com/radekd91/inferno
• VGGHead: https://github.com/KupynOrest/head_detector