SurFhead: Affine Rig Blending for Geometrically Accurate 2D Gaussian Surfel Head Avatars

ICLR 2025 Accepted

Installation

We heavily followed GaussianAvatars' instruction

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

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

git clone https://github.com/surfhead2025/SurFhead.git --recursive
cd SurFhead

conda create --name surfhead -y python=3.10
conda activate surfhead

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

Step 2: Setup paths (for Linux)

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

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. We provide the pre-processed data with this OneDrive link. Please request here to get access and download it into data/.

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

In the each shell, we curated all ablation studies. Last paragraph is our final version, SurFhead.

Training

To run the optimizer, simply use

sh train_cluster_external.sh

Test (self-reenactment)

To run the optimizer, simply use

sh test_cluster_external.sh

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.