NeurIPS 2025 [Project Page | arXiv]

Install the conda environment of ZeroGS.

conda create -n zero_gs python=3.9
conda activate zero_gs
cd ZeroGS/scripts
./scripts/env/install.sh

git hook for code style checking:

pre-commit install --hook-type pre-commit

• Release ACE0 implementation
  • Incorporate GLACE into ACE0
• Release our customized 3D Gaussian Splatting module
  • Incorporate Scaffold-GS
  • Incorporate DOGS
• Release ZeroGS implementation

We aim at providing a framework which makes it easy to implement your own neural implicit module with this codebase and since this project starts before the code releasing of ACE0, we re-implement ACE0 based on our codebase.

Before training ACE0, please download the pretrained feature encoder from ACE, and put it under the folder ZeroGS/conerf/model/scene_regressor.

conda activate zero_gs
visdom -port=9000 # Keep the port the same as the `visdom_port` provided in the configuration file
cd ZeroGS/scripts/train
./train_ace_zero.sh 0 ace_early_stop_resize_2k_anneal mipnerf360 ace

We use visdom to visualize the camera pose predictions during training. You can access https://localhost:9000 to view it.

conda activate zero_gs
cd ZeroGS/scripts/eval
./eval_ace_zero.sh 0 ace_early_stop_resize_2k_anneal mipnerf360 ace

Metrics file and camera poses will be recorded in eval/val/ folder. Point clouds are recorded in the eval/val/ACE0_COLMAP (This folder also contains the model files in COLMAP formats) in .ply format.

All the parameters related to train ACE0 are provided the configuration file in config/ace/mipnerf360.yaml. Most of the parameters can be kept the same as in this configuration file. However, the parameters listed below need to be adjusted accordingly to obtain better performance:

trainer:
  # We can use less iterations for the `garden` scene (i.e. 2000).
  min_iterations_per_epoch: 5000

pose_estimator:
  # Change this to a larger threshold (3000) for the 'garden` scene of the mipnerf360 dataset.
  min_inlier_count: 2000 # minimum number of inlier correspondences when registering an image

A larger value in min_iterations_per_epoch can make the mapping more accurate, but also lead to longer training time.

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

@inproceedings{yuchen2024zerogs,
    title={ZeroGS: Training 3D Gaussian Splatting from Unposed Images},
    author={Yu Chen, Rolandos Alexandros Potamias, Evangelos Ververas, Jifei Song, Jiankang Deng, Gim Hee Lee},
    booktitle={arXiv},
    year={2024},
}

This work is built upon ACE, DUSt3R, and Spann3R. We sincerely thank all the authors for releasing their code.

Copyright © 2024, Chen Yu. All rights reserved. Please see the license file for terms.