GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians

(An archive of the official repo)

GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians

Licenses

Toyota Motor Europe NV/SA and its affiliated companies retain all intellectual property and proprietary rights in and to this software and related documentation. Any commercial use, reproduction, disclosure or distribution of this software and related documentation without an express license agreement from Toyota Motor Europe NV/SA is strictly prohibited.

This project uses Gaussian Splatting, which carries its original license. The GUI is inspired by INSTA. The mesh rendering operations are adapted from NVDiffRec and NVDiffRast.

This work is made available under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International.

Setup

Hardware Requirements

CUDA-ready GPU with Compute Capability 7.0+
11 GB VRAM (we used RTX 2080Ti)

Software Requirements

Conda (recommended for easy setup)
C++ Compiler for PyTorch extensions (we used Visual Studio for Windows, GCC for Linux)
CUDA SDK for PyTorch extensions, install after Visual Studio or GCC
C++ Compiler and CUDA SDK must be compatible
FFMPEG to create result videos

Additional python packages

RoMa (for rotation representations)
DearPyGUI (for viewer interface)
NVDiffRast (for mesh rendering in viewer)

Tested Platforms

PyTorch Version	CUDA version	Linux	Windows (VS2022)	Windows (VS2019)
2.0.1	11.7.1	Pass	Fail to compile	Pass
2.2.0	12.1.1	Pass	Pass	Pass

Installation

Our default installation method is based on Conda package and environment management:

Step 1: Clone the repo and install `cuda-toolkit` with `conda`

git clone https://github.com/ShenhanQian/GaussianAvatars.git --recursive
cd GaussianAvatars

conda create --name gaussian-avatars -y python=3.10
conda activate gaussian-avatars

# Install CUDA and ninja for compilation
conda install -c "nvidia/label/cuda-11.7.1" cuda-toolkit ninja  # use the right CUDA version

Step 2a: Setup paths (for Linux)

ln -s "$CONDA_PREFIX/lib" "$CONDA_PREFIX/lib64"  # to avoid error "/usr/bin/ld: cannot find -lcudart"

Step 2b: Setup environment variables (for Windows with PowerShell)

conda env config vars set CUDA_PATH="$env:CONDA_PREFIX"  

## Visual Studio 2022 (modify the version number `14.39.33519` accordingly)
conda env config vars set PATH="$env:CONDA_PREFIX\Script;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.39.33519\bin\Hostx64\x64;$env:PATH"
## or Visual Studio 2019 (modify the version number `14.29.30133` accordingly)
conda env config vars set PATH="$env:CONDA_PREFIX\Script;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\bin\HostX86\x86;$env:PATH" 

# re-activate the environment to make the above eonvironment variables effective
conda deactivate
conda activate gaussian-avatars

Step 2c: Setup environment variables (for Windows with Command Prompt)

conda env config vars set CUDA_PATH=%CONDA_PREFIX%

## Visual Studio 2022 (modify the version number `14.39.33519` accordingly)
conda env config vars set PATH="%CONDA_PREFIX%\Script;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.39.33519\bin\Hostx64\x64;%PATH%"
## or Visual Studio 2019 (modify the version number `14.29.30133` accordingly)
conda env config vars set PATH="%CONDA_PREFIX%\Script;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\bin\HostX86\x86;%PATH%"

# re-activate the environment to make the above eonvironment variables effective
conda deactivate
conda activate gaussian-avatars

Step 3: Install PyTorch and other packages

# Install PyTorch (make sure that the CUDA version matches with "Step 1")
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu117
# or
conda install pytorch torchvision pytorch-cuda=11.7 -c pytorch -c nvidia
# make sure torch.cuda.is_available() returns True

# Install the rest packages (can take a while to compile diff-gaussian-rasterization, simple-knn, and nvdiffrast)
pip install -r requirements.txt

Data

Preprocessed NeRSemble Dataset

We use 9 subjects from NeRSemble dataset in our paper. You can download the pre-processed data from LRZ or OneDrive.

Please request here to get approval. Please also request for the raw dataset here although you do not need to download it to run this repo.

FLAME Model

Our code and the pre-processed data relies on FLAME 2023. Downloaded assets from https://flame.is.tue.mpg.de/download.php and store them in below paths:

flame_model/assets/flame/flame2023.pkl # FLAME 2023 (versions w/ jaw rotation)
flame_model/assets/flame/FLAME_masks.pkl # FLAME Vertex Masks

It is possible to run our method with FLAME 2020 by download to flame_model/assets/flame/generic_model.pkl. The FLAME_MODEL_PATH in flame_model/flame.py needs to be updated accordingly. And the FLAME tracking results should also be based on FLAME 2020 in this case.

Running

Training

To run the optimizer, simply use

SUBJECT=306

python train.py \
-s data/UNION10_${SUBJECT}_EMO1234EXP234589_v16_DS2-0.5x_lmkSTAR_teethV3_SMOOTH_offsetS_whiteBg_maskBelowLine \
-m output/UNION10EMOEXP_${SUBJECT}_eval_600k \
--port 60000 --eval --white_background --bind_to_mesh

Command Line Arguments for train.py

--source_path / -s

Path to the source directory containing a COLMAP or Synthetic NeRF data set.

--model_path / -m

Path where the trained model should be stored (output/<random> by default).

--eval

Add this flag to use a training/val/test split for evaluation.

--bind_to_mesh

Add this flag to bind 3D Gaussians to a driving mesh, e.g., FLAME.

--resolution / -r

Specifies resolution of the loaded images before training. If provided 1, 2, 4 or 8, uses original, 1/2, 1/4 or 1/8 resolution, respectively. For all other values, rescales the width to the given number while maintaining image aspect. If not set and input image width exceeds 1.6K pixels, inputs are automatically rescaled to this target.

--data_device

Specifies where to put the source image data, cuda by default, recommended to use cpu if training on large/high-resolution dataset, will reduce VRAM consumption, but slightly slow down training. Thanks to HrsPythonix.

--white_background / -w

Add this flag to use white background instead of black (default), e.g., for evaluation of NeRF Synthetic dataset.

--sh_degree

Order of spherical harmonics to be used (no larger than 3). 3 by default.

--convert_SHs_python

Flag to make pipeline compute forward and backward of SHs with PyTorch instead of ours.

--convert_cov3D_python

Flag to make pipeline compute forward and backward of the 3D covariance with PyTorch instead of ours.

--debug

Enables debug mode if you experience erros. If the rasterizer fails, a dump file is created that you may forward to us in an issue so we can take a look.

--debug_from

Debugging is slow. You may specify an iteration (starting from 0) after which the above debugging becomes active.

--iterations

Number of total iterations to train for, 30_000 by default.

--ip

IP to start GUI server on, 127.0.0.1 by default.

--port

Port to use for GUI server, 60000 by default.

--test_iterations

Space-separated iterations at which the training script computes L1 and PSNR over test set, 7000 30000 by default.

--save_iterations

Space-separated iterations at which the training script saves the Gaussian model, 7000 30000 <iterations> by default.

--checkpoint_iterations

Space-separated iterations at which to store a checkpoint for continuing later, saved in the model directory.

--start_checkpoint

Path to a saved checkpoint to continue training from.

--quiet

Flag to omit any text written to standard out pipe.

--feature_lr

Spherical harmonics features learning rate, 0.0025 by default.

--opacity_lr

Opacity learning rate, 0.05 by default.

--scaling_lr

Scaling learning rate, 0.005 by default.

--rotation_lr

Rotation learning rate, 0.001 by default.

--position_lr_max_steps

Number of steps (from 0) where position learning rate goes from initial to final. 30_000 by default.

--position_lr_init

Initial 3D position learning rate, 0.00016 by default.

--position_lr_final

Final 3D position learning rate, 0.0000016 by default.

--position_lr_delay_mult

Position learning rate multiplier (cf. Plenoxels), 0.01 by default.

--densify_from_iter

Iteration where densification starts, 500 by default.

--densify_until_iter

Iteration where densification stops, 15_000 by default.

--densify_grad_threshold

Limit that decides if points should be densified based on 2D position gradient, 0.0002 by default.

--densification_interal

How frequently to densify, 100 (every 100 iterations) by default.

--opacity_reset_interval

How frequently to reset opacity, 3_000 by default.

--lambda_dssim

Influence of SSIM on total loss from 0 to 1, 0.2 by default.

--percent_dense

Percentage of scene extent (0--1) a point must exceed to be forcibly densified, 0.01 by default.

By default, the trained models use all available images in the dataset. To train them while withholding a validation set and a test set for evaluation, use the --eval flag.

A complete evaluation on the validation set (novel-view synthesis) and test set (self-reenactment) will be conducted every --interval iterations. You can check the metrics in the terminal or within Tensorboard. Although we only save a few images in Tensorboard, the metrics are computed on all images.

Rendering

python render.py -m <path to trained model> # Generate renderings

Render the validation set (novel-view synthesis):

SUBJECT=306

python render.py \
-m output/UNION10EMOEXP_${SUBJECT}_eval_600k \
--skip_train --skip_test

Render the test set (self-reenactment):

SUBJECT=306

python render.py \
-m output/UNION10EMOEXP_${SUBJECT}_eval_600k \
--skip_train --skip_val

Render the test set (self-reenactment) only in a front view:

SUBJECT=306

python render.py \
-m output/UNION10EMOEXP_${SUBJECT}_eval_600k \
--skip_train --skip_val \
--select_camera_id 8  # front view

Cross-identity reenactment with the FREE sequence of TGT_SUBJECT:

SUBJECT=306
TGT_SUBJECT=218

python render.py \
-t data/${TGT_SUBJECT}_FREE_v16_DS2-0.5x_lmkSTAR_teethV3_SMOOTH_offsetS_whiteBg_maskBelowLine \
-m output/UNION10EMOEXP_${SUBJECT}_eval_600k \
--select_camera_id 8  # front view

Cross-identity reenactment with 10 prescribed motion sequences of TGT_SUBJECT:

SUBJECT=306
TGT_SUBJECT=218

python render.py \
-t data/UNION10_${TGT_SUBJECT}_EMO1234EXP234589_v16_DS2-0.5x_lmkSTAR_teethV3_SMOOTH_offsetS_whiteBg_maskBelowLine \
-m output/UNION10EMOEXP_${SUBJECT}_eval_600k \
--select_camera_id 8  # front view

Command Line Arguments for render.py

--model_path / -m

Path to the trained model directory you want to create renderings for.

--skip_train

Flag to skip rendering the training set.

--skip_val

Flag to skip rendering the test set.

--skip_test

Flag to skip rendering the validation set.

--quiet

Flag to omit any text written to standard out pipe.

--select_camera_id

Only render from a specific camera id.

--target_path / -t

Path to the target directory containing a motion sequence for reenactment.

The below parameters will be read automatically from the model path, based on what was used for training. However, you may override them by providing them explicitly on the command line.

--source_path / -s

Path to the source directory containing a COLMAP or Synthetic NeRF data set.

--images / -i

Alternative subdirectory for COLMAP images (images by default).

--eval

Add this flag to use a MipNeRF360-style training/test split for evaluation.

--resolution / -r

Changes the resolution of the loaded images before training. If provided 1, 2, 4 or 8, uses original, 1/2, 1/4 or 1/8 resolution, respectively. For all other values, rescales the width to the given number while maintaining image aspect. 1 by default.

--white_background / -w

Add this flag to use white background instead of black (default), e.g., for evaluation of NeRF Synthetic dataset.

--convert_SHs_python

Flag to make pipeline render with computed SHs from PyTorch instead of ours.

--convert_cov3D_python

Flag to make pipeline render with computed 3D covariance from PyTorch instead of ours.

Interactive Viewers

We provide two interactive viewers for our method: remote and real-time. Our viewing solutions are based on DearPyGUI.

Running the Remote Viewer

During training, one can monitor the training progress with the remote viewer

python remote_viewer.py --port 60000

The remote viewer can slow down training a lot. You may want to close it or check "pause rendering" when not viewing.

The viewer could get frozen and disconnected the first time you enable "show mesh". You can try switching it on and off or simply wait for a second.

Running the Local Viewer

After training, one can load and render the optimized 3D Gaussians with the local viewer

SUBJECT=306
ITER=300000

python local_viewer.py \
--point_path output/UNION10EMOEXP_${SUBJECT}_eval_600k/point_cloud/iteration_${ITER}/point_cloud.ply

Command Line Arguments for local_viewer.py

--point_path

Path to the gaussian splatting file (ply)

--motion_path

Path to the motion file (npz)

Cite

If you find our paper or code useful in your research, please cite with the following BibTeX entry:

@article{qian2023gaussianavatars,
  title={GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians},
  author={Qian, Shenhan and Kirschstein, Tobias and Schoneveld, Liam and Davoli, Davide and Giebenhain, Simon and Nie{\ss}ner, Matthias},
  journal={arXiv preprint arXiv:2312.02069},
  year={2023}
}

Name		Name	Last commit message	Last commit date
Latest commit History 307 Commits
arguments		arguments
assets		assets
flame_model		flame_model
gaussian_renderer		gaussian_renderer
lpipsPyTorch		lpipsPyTorch
mesh_renderer		mesh_renderer
reference_tracker		reference_tracker
scene		scene
submodules		submodules
utils		utils
.gitignore		.gitignore
.gitmodules		.gitmodules
LICENSE.md		LICENSE.md
LICENSE_GS.md		LICENSE_GS.md
README.md		README.md
convert.py		convert.py
local_viewer.py		local_viewer.py
metrics.py		metrics.py
remote_viewer.py		remote_viewer.py
render.py		render.py
requirements.txt		requirements.txt
train.py		train.py

License

baoachun/GaussianAvatars

Folders and files

Latest commit

History

Repository files navigation

GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians

Licenses

Setup

Hardware Requirements

Software Requirements

Additional python packages

Tested Platforms

Installation

Step 1: Clone the repo and install cuda-toolkit with conda

Step 2a: Setup paths (for Linux)

Step 2b: Setup environment variables (for Windows with PowerShell)

Step 2c: Setup environment variables (for Windows with Command Prompt)

Step 3: Install PyTorch and other packages

Data

Preprocessed NeRSemble Dataset

FLAME Model

Running

Training

--source_path / -s

--model_path / -m

--eval

--bind_to_mesh

--resolution / -r

--data_device

--white_background / -w

--sh_degree

--convert_SHs_python

--convert_cov3D_python

--debug

--debug_from

--iterations

--ip

--port

--test_iterations

--save_iterations

--checkpoint_iterations

--start_checkpoint

--quiet

--feature_lr

--opacity_lr

--scaling_lr

--rotation_lr

--position_lr_max_steps

--position_lr_init

--position_lr_final

--position_lr_delay_mult

--densify_from_iter

--densify_until_iter

--densify_grad_threshold

--densification_interal

--opacity_reset_interval

--lambda_dssim

--percent_dense

Rendering

--model_path / -m

--skip_train

--skip_val

--skip_test

--quiet

--select_camera_id

--target_path / -t

--source_path / -s

--images / -i

--eval

--resolution / -r

--white_background / -w

--convert_SHs_python

--convert_cov3D_python

Interactive Viewers

Running the Remote Viewer

Running the Local Viewer

--point_path

--motion_path

Cite

Step 1: Clone the repo and install `cuda-toolkit` with `conda`

Packages