Alex Hanson*, Allen Tu*, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, Tom Goldstein

arXiv | Website

* indicates equal contribution

This repository is the official code for the paper "PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting". It is built on top of the 3D-GS and LightGaussian codebases. Fisher computation in the included submodule is built using code from FisherRF. If you make use of this code in your research, please consider citing these works in addition to ours.

Abstract: Recent advances in novel view synthesis have enabled real-time rendering speeds with high reconstruction accuracy. 3D Gaussian Splatting (3D-GS), a foundational point-based parametric 3D scene representation, models scenes as large sets of 3D Gaussians. However, complex scenes can consist of millions of Gaussians, resulting in high storage and memory requirements that limit the viability of 3D-GS on devices with limited resources. Current techniques for compressing these pretrained models by pruning Gaussians rely on combining heuristics to determine which Gaussians to remove. At high compression ratios, these pruned scenes suffer from heavy degradation of visual fidelity and loss of foreground details. In this paper, we propose a principled sensitivity pruning score that preserves visual fidelity and foreground details at significantly higher compression ratios than existing approaches. It is computed as a second-order approximation of the reconstruction error on the training views with respect to the spatial parameters of each Gaussian. Additionally, we propose a multi-round prune-refine pipeline that can be applied to any pretrained 3D-GS model without changing its training pipeline. After pruning 90% of Gaussians, a substantially higher percentage than previous methods, our PUP 3D-GS pipeline increases average rendering speed by 3.56× while retaining more salient foreground information and achieving higher image quality metrics than existing techniques on scenes from the Mip-NeRF 360, Tanks & Temples, and Deep Blending datasets.

Follow the setup instruction for the original 3D-GS codebase, including recursively cloning the repository submodules.

conda env create --file environment.yml
conda activate gaussian_splatting_pup

We created an additional submodule for CUDA Fisher computation: rasterization_and_pup_fisher. Ensure that it is also cloned and installed.

To prune and evaluate a scene, set the environment variables SCENE_DATA_PATH, SCENE_MODEL_PATH, and SCENE_NAME, then run:

bash scripts/full_pruning_pipeline.sh

• SCENE_DATA_PATH is the path to the COLMAP or NeRF Synthetic dataset.

• SCENE_MODEL_PATH is the path to the pretrained 3D-GS model.

• SCENE_NAME is unique name for the scene.

This script will create the ./experiments/<SCENE_NAME>_80 and ./experiments/<SCENE_NAME>_80_50 directories for one-round 80% and two-round 90% pruning, respectively. Then, it will save the prune-refined models inside the point_cloud/iteration_35000 subdirectories. It will then render the test views using the 3D-GS render.py script, which has been updated to also output the average FPS in fps.txt. Finally, the image quality metrics of the test renders will be computed and saved in results.json using the 3D-GS metrics.py script.

To create Fisher matrices of a scene, set the environment variables SCENE_DATA_PATH and SCENE_MODEL_PATH, then run:

bash scripts/create_fishers.sh

• SCENE_DATA_PATH is the path to the COLMAP or NeRF Synthetic dataset.

• SCENE_MODEL_PATH is the path to the pretrained 3D-GS model.

Fishers will be saved in the fisher-pool-<X>/fishers_xyz_scaling subdirectory of the pretrained model directory, where X is the training iteration.

@InProceedings{HansonTuPUP3DGS,
    author    = {Hanson, Alex and Tu, Allen and Singla, Vasu and Jayawardhana, Mayuka and Zwicker, Matthias and Goldstein, Tom},
    title     = {PUP 3D-GS: Principled Uncertainty Pruning for 3D Gaussian Splatting},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {5949-5958},
    url       = {https://pup3dgs.github.io/}
}

For additional papers on efficient 3D Gaussian Splatting, see our group’s related work below. If your research builds on ours, we encourage you to cite these papers.

1. Speedy-Splat (CVPR 2025) [BibTeX] — Accelerate 3D Gaussian Splatting rendering speed by over 6× and reduce model size by over 90% through accurately localizing primitives during rasterization and pruning the scene during training, providing a significantly higher speedup than existing techniques while maintaining competitive image quality.
2. SpeeDe3DGS [BibTeX] — Boost DeformableGS rendering speed from 20 to 276 FPS using temporal sensitivity pruning and groupwise SE(3) motion distillation, all while preserving the superior image quality of per-Gaussian neural motion.

This research is based upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via IARPA R&D Contract No. 140D0423C0076. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Additional support was provided by ONR MURI program and the AFOSR MURI program. Commercial support was provided by Capital One Bank, the Amazon Research Award program, and Open Philanthropy. Zwicker was additionally supported by the National Science Foundation (IIS-2126407). Goldstein was additionally supported by the National Science Foundation (IIS-2212182) and by the NSF TRAILS Institute (2229885).