[Project Page] [arXiv] [Dataset]

LL-Gaussian is a novel framework for low-light scene reconstruction and novel view synthesis. It leverages 3D Gaussian Splatting with specialized modules to enhance rendering quality in extreme lighting conditions.

Hao Sun^1,2, Fenggen Yu⁴, Huiyao Xu³, Tao Zhang⁵, Changqing Zou^1,3
1Zhejiang Lab ²University of Chinese Academy of Sciences ³State Key Lab of CAD&CG, Zhejiang University
⁴Simon Fraser University ⁵Hangzhou Dianzi University

• Provide LLRS dataset.
• Provide inference code.
• Provide training code.

• 🌙 Low-Light Gaussian Initialization: Robust 3D Gaussian initialization without reliance on SfM tools like COLMAP.
• 🔀 Gaussian Decomposition: Dual-branch modeling of intrinsic vs. transient scene components.
• 🎥 Novel View Synthesis in the Dark: State-of-the-art rendering quality under low-light and nighttime conditions.

The code will be released soon.

If you find our work useful, please consider citing us:

@inproceedings{sun2025ll,
  title={LL-Gaussian: Low-Light Scene Reconstruction and Enhancement via Gaussian Splatting for Novel View Synthesis},
  author={Sun, Hao and Yu, Fenggen and Xu, Huiyao and Zhang, Tao and Zou, Changqing},
  booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
  pages={4261--4270},
  year={2025}
}

@article{sun2025ll,
  author={Sun, Hao and Yu, Fenggen and Xu, Huiyao and Zhang, Tao and Zou, Changqing},
  title={LL-Gaussian: Low-Light Scene Reconstruction and Enhancement via Gaussian Splatting for Novel View Synthesis},
  journal={arXiv preprint arXiv:2504.10331},
  year={2025},
}