FlexGaussian: Flexible and Cost-Effective Training-Free Compression for 3D Gaussian Splatting

Overview

FlexGaussian is a lightweight compression pipeline for 3D Gaussian Splatting (3DGS) that:

• Requires no training or fine-tuning
• Supports adjustable compression ratios
• Runs efficiently on both desktop and mobile hardware

Key features

• Training-Free: No retraining or fine-tuning required. Just plug and play.
• Cost-Effective: Compresses models in under 1 minute on desktops and under 3 minutes on mobile platforms.
• Flexible: With importance scores computed (a one-time cost), switching compression ratios takes 1–2 seconds.

Installation

We tested on both desktop and mobile platforms using Ubuntu 20.04, CUDA 11.6, PyTorch 1.12.1, and Python 3.7.13.

Install dependencies

# clone the repo
$ git clone [email protected]:Supercomputing-System-AI-Lab/FlexGaussian.git
$ cd FlexGaussian

# setup conda environment
$ conda env create --file environment.yml
$ conda activate FlexGaussian

Setup data paths

• open scripts/run.sh
• specify:
  • source_path: path to the dataset
  • model_path: path to the pre-trained 3DGS model
  • imp_score_path (optional): path to precomputed importance score in .npz format.

Data preparation

Datasets

• We have tested on Mip-NeRF360, DeepBlending, and Tanks&Temples.
• The Mip-NeRF360 scenes are hosted by the paper authors here.
• We use the Tanks&Temples and Deep Blending scenes hosted by the paper authors of 3DGS here.
• Note that while both training and test views are loaded, importance score computation does not rely on image content. It evaluates each 3D Gaussian's contribution by counting how many pixels it overlaps with across a view set (we use training views for simplicity, but any camera poses would work).
• The render resolution should match the training resolution.

Pre-trained 3DGS model

• To ensure consistency and reproducibility, we use pre-trained 3DGS models released by the official 3DGS repository. Pre-trained Models (14 GB)
• Other models compatible with the standard 3DGS format are also supported.

Usage

Quick start

To run the full compression pipeline with default settings:

$ bash scripts/run.sh [scene_name]

• scene_name can be one of the supported scenes from Mip-NeRF360, DeepBlending, or Tanks&Temples.
• Omit scene_name to run compression across all 13 scenes evaluated in the paper.

CLI usage

$ python compress.py -s [source_path] -m [model_path] --output_path [output_path] --eval

Optional arguments

• --quality_target_diff [float]
  Target PSNR drop. Default: 1.0 dB

• --imp_score_path [string]
  Path to precomputed importance score (skip score computation)

• --save_render
  Save rendered images from the compressed 3DGS model

Sample output

• Below is a an example output after compressing the bicycle scene:
• Compressed file is compared against input 3DGS model in quality and file size.
• The pipeline time is break down to I/O, importance score calculation, and search.

[INFO] Running on scene: bicycle
============================================================
Best combination found: pruning=0.4, sh=0.5

[INFO] Best config:           pruning_rate = 0.4, sh_rate = 0.5
[INFO] Base PSNR:             25.1758 dB
[INFO] Best PSNR:             24.1808 dB
[INFO] PSNR drop:             0.9950 dB
[INFO] Input file size:       1450.28 MB
[INFO] Output file size:      70.97 MB
[INFO] Compression ratio:     20.43x
============================================================

============================================================
[STATS] Pipeline Time Breakdown:
  Load time:      9.01 s
  Score time:     2.97 s
  Search time:   10.31 s
  Store time:     0.24 s
  Total time:    22.52 s
============================================================

Test devices

We tested on both desktop and mobile platforms using Ubuntu 20.04, CUDA 11.6, PyTorch 1.12.1, and Python 3.7.13. *Performance may vary based on the library versions and sources of precompiled binaries.

• Desktop:
  • GPU: Nvidia RTX 3090 (10496-cores Ampere arch)
  • CPU: Intel Core i9-10900K (10-cores)
  • RAM: 64 GB
• Mobile (Nvidia Jetson AGX Xavier):
  • GPU: GV10B (512-cores Volta arch)
  • CPU: Nvidia Carmel ARM v8 (8-cores)
  • RAM: 16 GB Unified Memory (shared CPU/GPU)

BibTeX

@inproceedings{FlexGaussian,
        title={FlexGaussian: Flexible and Cost-Effective Training-Free Compression for 3D Gaussian Splatting},
        author={Tian, Boyuan and Gao, Qizhe and Xianyu, Siran and Cui, Xiaotong and Zhang, Minjia},
        booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
        year={2025}

Acknowledgement

• Thanks to the authors of 3D Gaussian Splatting for their foundational work.
• We also acknowledge the authors of LightGaussian for their method of calculating importance scores.