Metropolis-Hastings Sampling for 3D Gaussian Reconstruction

NeurIPS 2025

Hyunjin Kim¹ · Haebeom Jung² · Jaesik Park²

¹UC San Diego · ²Seoul National University

News

• [11/26/2025] Code is now available!
• [09/25/2025] MH-3DGS is accepted to NeurIPS 2025 🎉

Setup

git clone https://github.com/hjhyunjinkim/MH-3DGS.git --recursive
cd MH-3DGS
conda config --set channel_priority flexible
conda env create --file environment.yml
conda activate mh_3dgs

Datasets

We use the same datasets used by 3DGS, the datasets can be found at:

• Mip-NeRF360
• Tanks&Temples
• Deep Blending

Running

To train, use the following command. We withhold a test set for evaluation by using the --eval flag.

python train.py -s <path to COLMAP or NeRF Synthetic dataset> --eval

For evaluation, use the following commands.

python render.py -m <path to trained model> # Generate renderings
python metrics.py -m <path to trained model> # Compute error metrics on renderings

You may also use the full_eval.py script for full evaluation of the datasets:

python full_eval.py -m360 <mipnerf360 folder> -tat <tanks and temples folder> -db <deep blending folder> --output_path <output folder>

Acknowledgements

Our work is based on the open-sourced official implementations of 3DGS and 3DGS-MCMC.

Citation

If you find our work useful, please consider citing:

@inproceedings{
  kim2025metropolishastings,
  title={Metropolis-Hastings Sampling for 3D Gaussian Reconstruction},
  author={Hyunjin Kim and Haebeom Jung and Jaesik Park},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025}
}