CoherentGS tackles one of the hardest regimes for 3D Gaussian Splatting (3DGS): Sparse inputs with severe motion blur. We break the "vicious cycle" between missing viewpoints and degraded photometry by coupling a physics-aware deblurring prior with diffusion-driven geometry completion, enabling coherent, high-frequency reconstructions from as few as 3–9 views on both synthetic and real scenes.

• 2025-11-28: Project page released at https://github.com/PotatoBigRoom/CoherentGS.
• Initial code, data links, and training scripts are now public.

Tested with Python 3.10 and PyTorch 2.1.2 (CUDA 11.8). Adjust CUDA wheels as needed for your platform.

# (Optional) fresh conda env
conda create --name CoherentGS -y "python<3.11"
conda activate CoherentGS

# Install dependencies
pip install --upgrade pip setuptools
pip install "torch==2.1.2+cu118" "torchvision==0.16.2+cu118" --extra-index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

• DL3DV-Blur and related assets: https://huggingface.co/datasets/Passwerob/CoherentGS/tree/main
• Additional samples and preprocessed splits: https://huggingface.co/datasets/Passwerob/CoherentGS/tree/main
  Place downloaded data under datasets/ (or adjust paths in the provided scripts).

Train on DL3DV-Blur (full resolution) with:

bash run_dl3dv.sh

For custom settings, start from run.sh and tweak dataset paths, resolution, and batch sizes.

If CoherentGS supports your research, please cite:

@article{feng2025coherentgs,
  author    = {Feng, Chaoran and Xu, Zhankuo and Li, Yingtao and Zhao, Jianbin and Yang, Jiashu and Yu, Wangbo and Yuan, Li and Tian, Yonghong},
  title     = {Breaking the Vicious Cycle: Coherent 3D Gaussian Splatting from Sparse and Motion-Blurred Views},
  year      = {2025},
}

This repo builds on outstanding open-source efforts, in particular gaussian-splatting, BAD-Gaussians, DIFIX3D+. We thank the original authors for making their code available.

For issues or reproducibility questions, please open a GitHub issue. PRs that improve robustness, data loading, or evaluation scripts are welcome.