BakedAvatar takes monocular video recordings of a person and produces a mesh-based representation for real-time 4D head avatar synthesis on various devices including mobiles.
![](./assets/teaser.jpg)
First, clone this repo:
git clone https://github.com/buaavrcg/BakedAvatar
cd BakedAvatar
Then, install the required environment. We recommend using Anaconda to manage your python environment. You can setup the required environment by the following commands:
conda env create -f environment.yml
conda activate BakedAvatar
Or you can setup the required environment manually:
conda create -n BakedAvatar python=3.10
conda activate BakedAvatar
# Install Pytorch (or follow specific instructions for your GPU on https://pytorch.org/get-started/locally/)
conda install pytorch torchvision pytorch-cuda=11.8 -c pytorch -c nvidia
# Install various required libraries
pip install accelerate configargparse chumpy opencv-python pymeshlab trimesh scikit-image xatlas matplotlib tensorboard tqdm torchmetrics face-alignment
# Install Pytorch3D (Or follow instructions in https://github.com/facebookresearch/pytorch3d/blob/main/INSTALL.md)
conda install pytorch3d -c pytorch3d # Linux only
# Install nvdiffrast
git clone https://github.com/NVlabs/nvdiffrast
cd nvdiffrast && pip install . && cd ..
# Install mise (for levelset extraction)
pip install Cython && pip install code/utils/libmise/
Finally, download FLAME model, choose FLAME 2020 and unzip it, copy 'generic_model.pkl' into ./code/flame/FLAME2020
We use the same data format as in IMavatar and PointAvatar.
You can download a preprocessed dataset from subject 1, subject 2, then unzip the files into data/datasets
folder. You should be able to see the paths of one subject's videos structured like data/datasets/<subject_name>/<video_name>
.
To generate your own dataset, please follow the instructions in the IMavatar repo.
# Configure your training (use DDP?), see https://huggingface.co/docs/accelerate for details
accelerate config
cd code
accelerate launch scripts/runner.py -c config/subject1.yaml -t train
# Extract the meshes
accelerate launch scripts/runner.py -c config/subject1.yaml -t mesh_export
# Precompute the textures and export MLP weights
# Note: change the path to the actual mesh_data.pkl path if you use a different config
accelerate launch scripts/runner.py -c config/subject1.yaml -t texture_export --mesh_data_path ../data/experiments/subject1/mesh_export/iter_30000/marching_cube/res_init16_up5/mesh_data.pkl
# Fine-tune the textures with higher resolution (512x512)
accelerate launch scripts/runner.py -c config/subject1.yaml -t fine_tuning --img_res 512 512 --batch_size 1 --mesh_data_path ../data/experiments/subject1/mesh_export/iter_30000/marching_cube/res_init16_up5/mesh_data.pkl
# evaluate fine-tuned meshes (result of stage-3)
accelerate launch scripts/runner.py -c config/subject1.yaml -t test --img_res 512 512 --use_finetune_model --mesh_data_path ../data/experiments/subject1/finetune_mesh_data/iter_30000/mesh_data.pkl
# evaluate baked meshes (result of stage-2)
accelerate launch scripts/runner.py -c config/subject1.yaml -t test --img_res 512 512 --use_finetune_model --mesh_data_path ../data/experiments/subject1/mesh_export/iter_30000/marching_cube/res_init16_up5/mesh_data.pkl
# evaluate implicit fields (result of stage-1)
accelerate launch scripts/runner.py -c config/subject1.yaml -t test --img_res 512 512
# run cross-identity reenactment in PyTorch code
# you may replace the reenact_data_dir with the path to the reenactment dataset and
# replace the reenact_subdirs with the subdirectories names
accelerate launch scripts/runner.py -c config/subject1.yaml -t test --img_res 512 512 --mesh_data_path ../data/experiments/subject1/finetune_mesh_data/iter_30000/mesh_data.pkl --reenact_data_dir ../data/datasets/soubhik --reenact_subdirs test
# export baked meshes and textures for the web demo
python scripts/unpack_pkl.py ../data/experiments/subject1/finetune_mesh_data/iter_30000/mesh_data.pkl --output ./mesh_data
# export the FLAME parameter sequence for reenactment
# The flame_params.json are from the files in the train and test subfolders of the dataset (e.g., ../data/datasets/soubhik/train/flame_params.json)
# You may export the sequences from the same identity for self-reenactment, or from different identities for cross-identity reenactment.
python scripts/export_flame_sequence.py <path to 1st flame_params.json> <path to 2nd flame_params.json> ... --output ./sequence_data
Copy the exported mesh_data
directory and sequence_data
directory into the root of the web demo and start the server.
The diectionary structure should be like:
web_demo
├── jsUtils
├── mesh_data
├── sequence_data
└── src
Make sure that you have installed Npm and Node.js, then run the following commands:
cd web_demo
npm install
npm run build
npm install --global serve
serve
Then, open your browser and visit http://localhost:8080/
. To run the real-time reenactment, you can select one of the buttons with the name of the sequence in the web demo.
If you find our code or paper useful, please cite as:
@article{bakedavatar,
author = {Duan, Hao-Bin and Wang, Miao and Shi, Jin-Chuan and Chen, Xu-Chuan and Cao, Yan-Pei},
title = {BakedAvatar: Baking Neural Fields for Real-Time Head Avatar Synthesis},
year = {2023},
issue_date = {December 2023},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3618399},
doi = {10.1145/3618399},
volume = {42},
number = {6},
journal = {ACM Trans. Graph.},
month = {sep},
articleno = {225},
numpages = {14}
}