NanoGS: Training-Free Gaussian Splat Simplification

---

NanoGS achieves substantial simplification ratios while preserving fidelity and geometry—without post-optimization or calibrated images.

How to use

Run the simplifier on a 3DGS .ply file:

python simplification.py --ply <input.ply> [-o <output.ply>] [-r RATIO] [--k K] [--merge_cap MERGE_CAP] [--opacity_threshold OPACITY_THRESHOLD] [--lam_geo LAM_GEO] [--lam_sh LAM_SH]

Required

Option	Description
--ply PLY	Input PLY file (3DGS ply file).

Optional

Option	Default	Description
-o, --output	auto from input + ratio	Output PLY path.
-r, --ratio	0.5	Fraction of splats to keep, in (0,1). Example: 0.25 keeps 25%.
--k	16	k for KNN candidates.
--merge_cap	0.5	Max merges per pass as ratio of original splat count (0.01–0.5).
--opacity_threshold	0.1	Prune splats with opacity below this before merging. The effective threshold is min(threshold, median(opacity)).
--lam_geo	1.0	Geometry term weight in merge cost.
--lam_sh	1.0	Spherical-harmonics term weight in merge cost.

Examples

# Keep 25% of splats, output auto-named (e.g. scene_0.25.ply)
python simplification.py --ply scene.ply -r 0.25

# Custom output and all optional parameters
python simplification.py --ply scene.ply -o simplified.ply -r 0.5 --k 16 --merge_cap 0.3 --opacity_threshold 0.1 --lam_geo 1.0 --lam_sh 1.0

Citation

If you find our code or paper helps, please consider giving us a star or citing:

@misc{xiong2026nanogstrainingfreegaussiansplat,
    title={NanoGS: Training-Free Gaussian Splat Simplification}, 
    author={Butian Xiong and Rong Liu and Tiantian Zhou and Meida Chen and Zhiwen Fan and Andrew Feng},
    year={2026},
    eprint={2603.16103},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2603.16103}, 
}
}

License

See LICENSE in this repository.