SizeGS: Size-aware Compression of 3D Gaussian Splatting via Mixed Integer Programming
Shuzhao Xie ∗, Jiahang Liu ∗, Weixiang Zhang, Shijia Ge, Sicheng Pan, Chen Tang, Yunpeng Bai, Cong Zhang, Xiaoyi Fan, Zhi Wang † · 2024
Read SizeGS: Size-aware Compression of 3D Gaussian Splatting via Mixed Integer Programming by Shuzhao Xie ∗, Jiahang Liu ∗, Weixiang Zhang, Shijia Ge, Sicheng Pan, Chen Tang, Yunpeng Bai, Cong Zhang, Xiaoyi Fan, Zhi Wang † (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-05808.txt.
Introduction
Xie and co-authors present SizeGS, a method for squeezing 3D Gaussian splatting scenes to a target file size. The format reconstructs 3D scenes well but produces heavy data that strains storage and networks, and most compression tools cannot hit a chosen size to match changing bandwidth. The team measured which settings control size most: the share of Gaussians kept and the bit widths of their attributes.
They cast the choice of settings as a mixed-integer optimization problem. The goal is to maximize visual quality under a size budget. They then split the problem into a sampling step and an integer programming step, sped up by learned estimators of quality loss and size. On several scene variants the method led post-training compression methods and, after fine-tuning, rivaled tools that need full retraining. The catalog byline is garbled with footnote marks; the text names ten authors.
Catalog record
| CREATOR | Shuzhao Xie ∗, Jiahang Liu ∗, Weixiang Zhang, Shijia Ge, Sicheng Pan, Chen Tang, Yunpeng Bai, Cong Zhang, Xiaoyi Fan, Zhi Wang † |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-05808
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
SizeGS: Size-aware Compression of 3D Gaussian Splatting via Mixed Integer Programming Thanks: ∗ Shuzhao Xie and Jiahang Liu contributed equally to this research. Thanks: † Corresponding author Shuzhao Xie ∗, Jiahang Liu ∗, Weixiang Zhang, Shijia Ge, Sicheng Pan, Chen Tang, Yunpeng Bai, Cong Zhang, Xiaoyi Fan, Zhi Wang † Abstract . Recent advances in 3D Gaussian Splatting (3DGS) have greatly improved 3D reconstruction. However, its substantial data size poses a significant challenge for transmission and storage. While many compression techniques have been proposed, they fail to efficiently adapt to fluctuating network bandwidth, leading to resource wastage. We address this issue from the perspective of size-aware compression, where we aim to compress 3DGS to a desired size by quickly searching for suitable hyperparameters. Through a measurement study, we identify key hyperparameters that affect the size—namely, the reserve ratio of Gaussians and bit-width settings for Gaussian attributes. Then, we formulate this hyperparameter optimization problem as a mixed-integer nonlinear programming (MINLP) problem, with the goal of maximizing visual quality while respecting the size budget constraint. To solve the MINLP, we decouple this problem into two parts: discretely sampling the reserve ratio and determining the bit-width settings using integer linear programming (ILP). To solve the ILP more quickly and accurately, we design a quality loss estimator and a calibrated size estimator, as well as implement a CUDA kernel. Extensive experiments on multiple 3DGS variants demonstrate that our method achieves state-of-the-art performance in post-training compression. Furthermore, our method can achieve comparable quality to leading training-required methods after fine-tuning. Project page & Code: shuzhaoxie.github.io/sizegs . Conference: Proceedings of the 33rd ACM International Conference on Multimedia; October 27–31, 2025; Dublin, Ireland Proceedings of the 33rd ACM International Conference on Multimedia (MM ’25), October 27–31, 2025, Dublin, Ireland ISBN: 979-8-4007-2035-2/2025/10 DOI: 10.1145/3746027.3755370 CCS: Computing methodologies Image compression CCS: Computing methodologies Point-based models CCS: Computing methodologies Discrete space search CCS: Computing methodologies 3D imaging † † cc-license: by-nc-sa 1. Introduction In recent years, 3D Gaussian Splatting (3DGS) ( Kerbl et al., 2023 ) has revolutionized 3D scene representation and has been widely adopted in a variety of applications ( Wang et al., 2024b ; Li et al., 2024a ; Chen et al., 2025a ; Ziwen et al., 2024 ; Li et al., 2024b ; Li et al., 2024c ; Gu et al., 2024 ; Yao et al., 2025 ; Tan et al., 2024 ) . Despite its success, 3DGS still faces limitations in transmission efficiency due to its giant number of points and complex attributes. Though various 3DGS compression methods have been proposed ( Bagdasarian et al., 2024 ) , they overlook requirements arising from on-demand applications such as volumetric video streaming ( Wang et al., 2024b ; Zhang et al., 2022 ) , live streaming ( Guan et al., 2023 ) , and remote teleoperation ( Li et al., 2024a ) . These applications frequently encounter fluctuating network bandwidth, leading to jitter and blurriness that significantly impact user experience. These inefficiencies are further exacerbated in large-scale 3DGS scenes, which often exceed 20 megabytes (MB) even after compression. To overcome this challenge, size-aware 3DGS compression presents a promising direction. The key idea is to compress the 3DGS to desired size by automatically searching the suitable hyperparameter set . Based on the way of hyperparameter configuration, current 3DGS compression methods fall into two categories. One is the offline method ( Morgenstern et al., 2024 ; Papantonakis et al., 2024 ; Wu and Tuytelaars, 2024 ; Yang et al., 2024a ; Xie et al., 2024 ; Niedermayr et al., 2024 ; Fan et al., 2024 ) , which requires manually setting hyperparameters to adjust the compressed file size. Moreover, there are a variety of hyperparameters to consider, which demands a significant amount of human effort for adjustment. The other is the online method ( Lee et al., 2024 ; Liu et al., 2024c ; Chen et al., 2024b ; Wang et al., 2024c ; Wang et al., 2024a ) , which selects hyperparameters based on a context model and imports an extra hyperparameter \lambda to balance rate and bit consumption. While \lambda can adjust the final size, each adjustment requires retraining the context model from scratch, which is time-consuming. Additionally, it is difficult to predict the final size given a certain \lambda . Therefore, existing methods are unable to search for feasible hyperparameters under the constraint of the size budget in a short amount of time. Given the fast speed of recent integer programming solvers ( Yao et al., 2021 ) , we propose to achieve fast size-aware compression via mixed integer programming. However, searching for appropriate 3DGS compression hyperparameters from an optimization perspective is non-trivial. First, 3DGS compression pipeline has lots hyperparameters. Some hyperparameters have a significant impact on size, while others are not. It is necessary to filter out a set of hyperparameters that significantly influence size. Second, to formulate an integer programming model, an explicit analytical relationship needs to be established between size and hyperparameters, as well as between quality and hyperparameters. This model can then use existing solvers to determine the appropriate hyperparameter settings. Finally, to fastly achieve the accurate results, an accurate approximation function for the actual size and acceleration techniques should be propose, which will allow the solver determine whether a candidate hyperparameter solution satisfies the constraints faster. To tackle these challenges, we propose SizeGS , a codec that can compression 3DGS to a desired file size while maximizing visual quality. First, we summarize and analyze the pipeline of 3DGS compression and identify the two most influential hyperparameters on size: the reserve ratio \tau and bit-width setting {\mathbf{Q}} . Then, we formulate the problem as a Mixed Integer Nonlinear Programming (MINLP) model, where the objective is visual quality and the constraint is size. To solve it, we decompose the MINLP into two subproblems: one is discrete sampling for \tau , and the other is Integer Linear Programming (ILP) for {\mathbf{Q}} , which can be quickly solved using existing solvers ( Roy and Mitchell, 2020 ) . Specifically, we perform multiple rounds of search to find suitable hyperparameters. In each round, we sample a value for \tau , solve for {\mathbf{Q}} using ILP, compute the quality. We select the hyperparameters with the highest quality as the search result. To ensure that the bit-width settings search aligns with the ILP definition (i.e., both objective and constraint are linear functions of {\mathbf{Q}} ), we propose using quantization loss to replace visual quality and construct a linear analytical relationship between size and {\mathbf{Q}} for accurate size estimation. To accelerate the solution process, we implement a CUDA kernel for parallel quantization and reuse the search results from the previous round as the solver’s search starting point. Finally, we carefully design a three-step fine-tuning method to improve accuracy. In summary, our contributions are as follows: 1) We formulate the hyperparameter search problem in 3DGS compression as a Mixed Integer Nonlinear Programming (MINLP) model. To make this optimization tractable, we decouple this problem into two subproblems: discrete sampling for the reserve ratio and Integer Linear Programming (ILP) for bit-width setting. 2) To enable precise size estimation, we establish an analytical linear relationship between the bit-width setting and the resulting file size.