Xuangeng Chu1
Tatsuya Harada1,2
1The University of Tokyo,
2RIKEN AIP
GAGAvatar achieves one-shot 3DGS-based head reconstruction and ⚡️real-time⚡️ reenactment.
🔥 More results can be found in our Project Page. 🔥
git clone --recurse-submodules [email protected]:xg-chu/GAGAvatar.git
cd GAGAvatar
conda env create -f environment.yml
conda activate GAGAvatar
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 [email protected]:xg-chu/diff-gaussian-rasterization.git
pip install ./diff-gaussian-rasterization
rm -rf ./diff-gaussian-rasterization
Prepare resources with bash ./build_resources.sh.
Also prepare resources for GAGAvatar_track using:
cd core/libs/GAGAvatar_track
bash ./build_resources.sh
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.
You can use the pre-trained model directly, but if you need to retrain on your data:
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
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'])
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"]
}
python train.py --config gaga --dataset vfhq
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}
}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.
- 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
