GGHead: Fast and Generalizable 3D Gaussian Heads

Paper | Video | Project Page

Tobias Kirschstein, Simon Giebenhain, Jiapeng Tang, Markos Georgopoulos, and Matthias Nießner
Siggraph Asia 2024

1. Setup

1.1. Dependencies

1. Create conda environment gghead with newest PyTorch and CUDA 11.8:

   conda env create -f environment.yml

2. Ensure that nvcc.exe is taken from the conda environment and includes can be found:
   1. [Linux]

      conda activate gghead
      conda env config vars set CUDA_HOME=$CONDA_PREFIX
      conda activate base
      conda activate gghead

   2. [Windows]

      conda activate gghead
      conda env config vars set CUDA_HOME=$Env:CONDA_PREFIX
      conda env config vars set NVCC_PREPEND_FLAGS="-I$Env:CONDA_PREFIX\Library\include"
      conda activate base
      conda activate gghead

3. Check whether the correct nvcc can be found on the path via:

   nvcc --version

   which should say something like release 11.8.
4. Install Gaussian Splatting (which upon installation will compile CUDA kernels with nvcc):

   pip install gaussian_splatting@git+https://github.com/tobias-kirschstein/gaussian-splatting.git

   1. [Optional] If you compile the CUDA kernels on a different machine than the one you use for running code, you may need to manually specify the target GPU compute architecture for the compilation process via the TORCH_CUDA_ARCH_LIST environment variable:

      TORCH_CUDA_ARCH_LIST="8.0" pip install gaussian_splatting@git+https://github.com/tobias-kirschstein/gaussian-splatting.git

      Choose the correct compute architecture(s) that match your setup. Consult this website if unsure about the compute architecture of your graphics card.
   2. [Troubleshooting] On a Linux machine, if you run into

      gcc: fatal error: cannot execute 'cc1plus': posix_spawnp: No such file or directory
      

      or

      x86_64-conda_cos6-linux-gnu-cc: error trying to exec 'cc1plus': execvp: No such file or directory
      

      try

      conda install gxx_linux-64 gcc_linux-64
      

5. Finally install the gghead module via:

   pip install -e .

1.2. Environment Paths

All paths to data / models / renderings are defined by environment variables.
Please create a file in your home directory in ~/.config/gghead/.env with the following content:

GGHEAD_DATA_PATH = "..."
GGHEAD_MODELS_PATH = "..."
GGHEAD_RENDERINGS_PATH = "..."

Replace the ... with the locations where data / models / renderings should be located on your machine.

• GGHEAD_DATA_PATH: Location of the FFHQ dataset and foreground masks. Only needed for training. See Section 2 for how to obtain the datasets.
• GGHEAD_MODELS_PATH: During training, model checkpoints and configs will be saved here. See Section 4 for downloading pre-trained models.
• GGHEAD_RENDERINGS_PATH: Video renderings of trained models will be stored here

If you do not like creating a config file in your home directory, you can instead hard-code the paths in the env.py.

2. Data

Only needed for training. Download the zip files of the respective datasets and put them into ${GGHEAD_DATA_PATH}.

Dataset	Images + Cameras	Masks
FFHQ-512	FFHQ_png_512.zip	FFHQ_png_512_masks_modnet.zip
FFHQ-1024	FFHQ_png_1024.zip	FFHQ_png_1024_masks_modnet.zip
AFHQ-512	afhq_v2_processed.zip	afhq_v2_processed_masks_modnet.zip

The .zip files for "Images + Cameras" were created with the dataset creation script of EG3D at the respective resolution. The .zip files for "Masks" were obtained by running background matting module MODNet on each image.

The dataset files are under a Creative Commons BY-NC-SA 4.0 license being derivatives of the FFHQ Dataset. This means, you can use, redistribute, and adapt it for non-commercial purposes, as long as you (a) give appropriate credit by citing the StyleGAN paper, (b) indicate any changes that you've made, and (c) distribute any derivative works under the same license (https://creativecommons.org/licenses/by-nc-sa/4.0/)

3. Usage

3.1. Training

3.1.1. FFHQ 512

Stage 1 [256x256 pre-training]

BW_IMPLEMENTATION=1 python scripts/train_gghead.py ffhq FFHQ_png_512.zip 1 32 --kimg 6400

will start training GGHead on 1 GPU with a batch-size of 32 for 6400k images. To speed up training, you can use more GPUs. E.g., python scripts/train_gghead.py ffhq FFHQ_png_512.zip 4 32 will train on 4 GPUs with a batch size of 8 per GPU instead.
Assets produced during training:

• Loss curves will be logged to weights&biases into a project called generative-gaussian-heads
• Generated images will be logged periodically to ${GGHEAD_MODELS_PATH}/gghead/GGHEAD-xxx every 200k train images
• Checkpoints are stored in ${GGHEAD_MODELS_PATH}/gghead/GGHEAD-xxx/checkpoints every 200k train images
• Evaluation result with FID scores are stored in ${GGHEAD_MODELS_PATH}/gghead/GGHEAD-xxx/evaluations every 200k train images

Stage 2 [512x512 training]

BW_IMPLEMENTATION=1 python scripts/train_gghead.py ffhq FFHQ_png_512.zip 1 32 --kimg 25000 --resume_run GGHEAD-xxx --overwrite_resolution 512 --overwrite_n_uniform_flame_vertices 512 --overwrite_lambda_tv_uv_rendering 100 --overwrite_lambda_beta_loss 1

Replace GGHEAD-xxx with the name of the run from stage 1, e.g., GGHEAD-4 (everything after the number can be omitted).

Useful flags

• use_vis_window: for local debugging. Opens a dearpygui window showing live training progress
• image_snapshot_ticks: How often debug images will be stored during training. Default: every 50 ticks = every 200k train images (1 tick = 4k images)
• metrics: Which FID scores to compute during training. Default: fid100,fid1k,fid10k. Always computing FID score with 50k generated samples is expensive during training. In our experience, only generating 10k images is already enough to assess which run performs better.

3.1.2. FFHQ 1024

BW_IMPLEMENTATION=1 python scripts/train_gghead.py ffhq FFHQ_png_1024.zip 1 32 --kimg 27000 --resume_run GGHEAD-xxx --overwrite_resolution 1024

Replace GGHEAD-xxx with the name of the run from stage 2.

3.1.3. AFHQ

BW_IMPLEMENTATION=1 python scripts/train_gghead.py ffhq afhq_v2_processed.zip 1 32 --kimg 30000 --resume_run GGHEAD-xxx

Replace GGHEAD-xxx with the name of the run from stage 2.

3.2. Rendering

3.2.1. Sampling 3D heads

From a trained model GGHEAD-xxx, render short videos of randomly sampled 3D heads via:

python scripts/sample_heads.py GGHEAD-xxx

Replace xxx with the actual ID of the model.
The generated videos will be placed into ${GGHEAD_RENDERINGS_PATH}/sampled_heads/

3.2.2. Interpolations

From a trained model GGHEAD-xxx, render interpolation videos that morph between randomly sampled 3D heads via:

python scripts/render_interpolation.py GGHEAD-xxx

Replace xxx with the actual ID of the model.
The generated videos will be placed into ${GGHEAD_RENDERINGS_PATH}/interpolations/

3.3. Evaluation

python scripts/evaluate_fid.py GGHEAD-xxx

Calculates FID score between generated images and the dataset images for the model GGHEAD-xxx (Replace xxx with the specific run ID that you want to evaluate).
The default number of generated samples for FID calculation is 50000 which can be changed via --fid.
The evaluation result will be printed in the terminal and also stored as a JSON file in ${GGHEAD_MODELS_PATH}/gghead/GGHEAD-xxx/evaluations.

3.5. Example Notebooks

The notebooks folder contains minimal examples on how to:

• Load a trained model, generate a 3D head and render it from an arbitrary viewpoint (inference.ipynb)

3.6. Visualizer

You can start the excellent GUI from EG3D and StyleGAN by running:

python visualizer.py

In the visualizer, you can select all checkpoints found in ${GGHEAD_MODELS_PATH}/gghead and freely explore the generated heads in 3D.

4. Downloads

4.1. Pre-trained models

Put pre-trained models into ${GGHEAD_MODELS_PATH}/gghead.

Dataset	GGHead model
FFHQ-512	GGHEAD-1_ffhq512
FFHQ-1024	GGHEAD-2_ffhq1024
AFHQ-512	GGHEAD-3-afhq512

---

@article{kirschstein2024gghead,
  title={GGHead: Fast and Generalizable 3D Gaussian Heads},
  author={Kirschstein, Tobias and Giebenhain, Simon and Tang, Jiapeng and Georgopoulos, Markos and Nie{\ss}ner, Matthias},
  journal={arXiv preprint arXiv:2406.09377},
  year={2024}
}

Contact Tobias Kirschstein for questions, comments and reporting bugs, or open a GitHub issue.