EnerGS

[ICML 2026] EnerGS: Energy-Based Gaussian Splatting with Partial Geometric Priors

Rui Song^1,2, Tianhui Cai¹, Markus Gross³, Yun Zhang¹, Walter Zimmer¹, Zhiyu Huang¹, Olaf Wysocki², Jiaqi Ma¹

¹University of California, Los Angeles    ²University of Cambridge    ³Technical University of Munich

Overview

EnerGS reformulates partial geometric supervision in 3D Gaussian Splatting as a continuous energy field that provides soft, observability-aware guidance during optimization. Instead of treating LiDAR (or any partial geometric prior) as a hard constraint, EnerGS partitions space into Occupied / Free / Unknown regions and induces an adaptive potential landscape over them, enabling robust reconstruction in LiDAR-blind but camera-visible areas while still rejecting floaters in certified free space.

Installation

1. Clone the repository

git clone --recursive https://github.com/ucla-mobility/EnerGS.git
cd EnerGS

2. Create the conda environment

Option A: Provided environment file (CUDA 12.1)

conda env create -f gs_cuda12.yaml
conda activate gs_cuda12

Option B: Manual setup

conda create -n energs python=3.10
conda activate energs

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install plyfile tqdm scipy numpy pillow lpips

3. Build the CUDA extensions

pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn
pip install submodules/fused-ssim  # optional, faster SSIM

Data Preparation

EnerGS expects:

1. A standard 3DGS-style COLMAP scene.
2. A pre-computed geometric field cache field_cache.npz (Euclidean Distance Transform + region labels derived from LiDAR ray-tracing). The toolkit that generates this cache will be released alongside the dataset bundle.

Expected layout:

<scene>/
├── images/
├── sparse/0/
│   ├── cameras.bin
│   ├── images.bin
│   └── points3D.bin
└── field_cache.npz   # geometric energy / gradient field

Training

python train_energs.py \
    -s <path/to/scene> \
    -m <output/path> \
    --field_npz <path/to/field_cache.npz> \
    --iterations 30000 \
    --eval

Key EnerGS-specific flags:

Flag	Default	Description
--energs_w_occ	1.0	Attractor weight in occupied space
--energs_sigma_occ	1.0	Welsch kernel bandwidth (m) in occupied space
--energs_w_unk	0.25	Weak attractor weight in unknown space
--energs_barrier_lambda	2.0	Boltzmann/Softplus barrier strength in free space
--energs_relax_lr	0.005	Geometric relax step size η_μ
--energs_prune_free	False	Enable discrete FREE-space pruning

Rendering and Evaluation

python render.py  -m <output/path>
python metrics.py -m <output/path>

Reported metrics include PSNR / SSIM (photometry) and Leak / OccCov / Margin / Thick (geometry), measured against a unified voxelized occupancy field derived from LiDAR.

Project Structure

EnerGS/
├── train_energs.py        # Main training entrypoint
├── render.py              # Novel-view rendering
├── metrics.py             # PSNR / SSIM / geometry metrics
├── energs/                # EnerGS core module
│   ├── geometric_energy.py    # E_occ / E_free / E_unk and gradient field
│   ├── gaussian_model.py      # GaussianModelEnerGS (decoupled update)
│   └── coverage.py            # Coverage modulation / FREE pruning
├── scene/                 # Scene loading (COLMAP + field cache)
├── gaussian_renderer/     # Differentiable rasterization wrappers
├── arguments/             # CLI argument definitions
└── submodules/            # CUDA extensions

Acknowledgments

EnerGS is built on top of 3D Gaussian Splatting by Kerbl et al. We thank the authors of 3DGS, 2DGS, Taming-3DGS, GeoGaussian, Mip-Splatting, Scaffold-GS, Street Gaussians, SplatAD, and the broader Gaussian Splatting community for releasing their code.

Citation

If you find this repository useful for your research, please consider giving us a star and citing our paper.

@inproceedings{song2026energs,
  title={EnerGS: Energy-Based Gaussian Splatting with Partial Geometric Priors},
  author={Song, Rui and Cai, Tianhui and Gross, Markus and Zhang, Yun and Zimmer, Walter and Huang, Zhiyu and Wysocki, Olaf and Ma, Jiaqi},
  booktitle={Proceedings of the International Conference on Machine Learning (ICML)},
  year={2026}
}