This repository contains a plug-and-play implementation of HyperGaussians, which can be directly installed from this repository for a frictionless integration into existing pipelines. We kindly refer to the original paper for a detailed description of the method and its applications.

For a real-world example that uses the HyperGaussians module, please refer to the flash_avatar branch of the repository. It contains our modified version of FlashAvatar with setup and usage instructions.

Install package using pip:

pip install git+https://github.com/gserifi/HyperGaussians.git

Note that the package relies on an existing PyTorch installation.

HyperGaussians can be imported and used just like any other torch.nn.Module. The constructor takes the number of Gaussians, the latent dimension, and the output dimension as arguments. The output dimension is typically 3 for position, 4 for rotation (quaternion), and 3 for scale. Depending on the application, the output values may need to be passed through certain activation functions to map them to the desired range or format.

import torch
from hypergaussians import HyperGaussians

# Initialization
num_gaussians = 10000
latent_dim = 8

hgs_xyz = HyperGaussians(num_gaussians, latent_dim, 3) # Position
hgs_rot = HyperGaussians(num_gaussians, latent_dim, 4) # Rotation
hgs_scl = HyperGaussians(num_gaussians, latent_dim, 3) # Scale

# Optionally specify initial output values
# initial_xyz = torch.randn(num_gaussians, 3)
# hgs_xyz = HyperGaussians(num_gaussians, latent_dim, 3, initial_output=initial_output)

# ...

# Conditioning

latent = torch.randn(1, num_gaussians, latent_dim)
deforms_xyz, uncertainty_xyz = hgs_xyz(latent)
deforms_rot, uncertainty_rot = hgs_rot(latent)
deforms_scl, uncertainty_scl = hgs_scl(latent)

# ... apply activation functions to restrict value range (e.g. using tanh or exp) or to convert to quaternions ...

If you find HyperGaussians useful for your research, please consider citing our paper:

@article{Serifi2025HyperGaussians,
  author={Gent Serifi and Marcel C. Bühler},
  title={HyperGaussians: High-Dimensional Gaussian Splatting for High-Fidelity Animatable Face Avatars},
  journal={arXiv},
  volume={2507.02803},
  year={2025}
}