GNF: Gaussian Neural Fields

📄 Paper | 🌐 Project Page |

This is the official implementation of GNF (Gaussian Neural Fields), featuring our novel RBF-based decoder for neural field modeling. The codebase supports various encoding methods and is designed for flexible, efficient neural field representation.

Main Contribution: RBF Decoder

Our key contribution is a novel Radial Basis Function (RBF) decoder that offers:

• Improved approximation quality through adaptive basis functions
• Compact representation with efficient parameter usage
• Fast training and inference
• Multiple basis function options (Gaussian, Multiquadric, Inverse Multiquadric, etc.)

Features

• Flexible encoding system supporting:
  • DiF-Grid
  • HashGrid
  • K-Planes
  • Hybrid K-Planes (K-Planes factorization with hash-based feature storage)
• Configurable decoders:
  • RBF (our main contribution)
  • MLP (baseline)
• Clean, minimal codebase for research and development

Getting Started

• Configuration files:
  • RGB training: configs/rgb.yaml
  • SDF training: configs/sdf.yaml
  • NeRF training: configs/nerf.yaml
• Main training scripts:
  • RGB training: main_rgb.py
  • SDF training: main_sdf.py
  • NeRF training: train_per_scene.py

Usage

# RGB Training
python main_rgb.py <image_file> --encoder <encoder_type> --decoder rbf

# SDF Training
python main_sdf.py <mesh_file> --encoder <encoder_type> --decoder rbf

# NeRF Training
python train_per_scene.py configs/nerf.yaml defaults.expname=<scene_name> dataset.datadir=<dataset_path>

Installation

Tested on Windows 11 and Ubuntu 20.04 with PyTorch 1.13.0

Install environment:

conda create -n gnf python=3.9
conda activate gnf
pip install -r requirements.txt 

Windows notes for the CUDA HashGrid encoder (optional)

Windows: CUDA HashGrid encoder

pip install .\gridencoder

If the wheel has to be compiled, just make sure your Visual Studio Build Tools use the same CUDA version that was used to compile your PyTorch wheel (see torch.version.cuda). Then:

cd gridencoder
pip install .

Citation

If you use this codebase, please cite our work.

@article{bouzidi2025GNF,
  title={GNF: Gaussian Neural Fields for Multidimensional Signal Representation and Reconstruction},
  author={Bouzidi, Abdelaziz and Laga, Hamid, Wannous, Hazem and Sohel Ferdous},
  journal={arXiv preprint arXiv:2503.06762},
  year={2025}
}

Acknowledgements

This implementation re-uses ideas and portions of code from the following open-source projects:

• autonomousvision/factor-fields – the original FactorFields.py implementation for radiance fields.
• ashawkey/torch-ngp – hash-grid CUDA encoder and training utilities.

We thank the authors for releasing their code under permissive licences, which made this work possible.