PyTorch implementation of "Self-Ensembling Gaussian Splatting for Few-Shot Novel View Synthesis", ICCV 2025, Oral.

# Create conda environment
conda create -n gaussian python=3.9
conda activate gaussian

# Install dependencies
bash setup.sh

We evaluate our method on the LLFF, DTU, Mip-NeRF360, and MVImgNet datasets. Note that, due to the stochastic nature of 3DGS, the evaluation results might be slightly different from those reported in the main paper.

1. Download LLFF from here.

2. Update base_path in tools/colmap_llff.py to the actual path of your data.

3. Run COLMAP to initialize point clouds and camera parameters:

   python tools/colmap_llff.py

bash scripts/run_llff.sh {your data path}

1. Download Mip-Nerf360 from here.

2. Update base_path in tools/colmap_360.py to the actual path of your data.

3. Run COLMAP to initialize point clouds and camera parameters:

   python tools/colmap_360.py

bash scripts/run_360.sh {your data path}

1. Follow the instructions here to download and organize the dataset. Download the masks from here

2. Update base_path in tools/colmap_dtu.py to the actual path of your data.

3. Run COLMAP to initialize point clouds and camera parameters:

   python tools/colmap_dtu.py

Note that COLMAP fails in some cases (scan8,scan40,scan110 with 3 views; scan21 with 6 views), so we use randomly initialized point clouds.

Update mask_path in copy_mask_dtu.sh accordingly, and then:

bash scripts/run_dtu.sh {your data path}

• Evaluation on LLFF
• Evaluation on Mip-Nerf360
• Evaluation on DTU
• Evaluation on MvImgNet

If you find the project useful, please consider citing:

@article{zhao2024self,
  title={Self-Ensembling Gaussian Splatting for Few-Shot Novel View Synthesis},
  author={Zhao, Chen and Wang, Xuan and Zhang, Tong and Javed, Saqib and Salzmann, Mathieu},
  journal={arXiv preprint arXiv:2411.00144},
  year={2024}
}