GS^2: Graph-based Spatial Distribution Optimization for Compact 3D Gaussian Splatting

Xianben Yang, Tao Wang, Yuxuan Li, Yi Jin, Haibin Ling · 2026

Read GS^2: Graph-based Spatial Distribution Optimization for Compact 3D Gaussian Splatting by Xianben Yang, Tao Wang, Yuxuan Li, Yi Jin, Haibin Ling (2026) 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-2604-01884.txt.

Introduction

Yang, Wang, Li, Jin and Ling cut the memory cost of 3D Gaussian Splatting, a way to rebuild scenes from soft 3D blobs for new viewpoints. Their 2026 method, GS squared, reshapes where blobs sit, adds points only where a statistical bound says they help, and prunes faint ones step by step. It keeps picture quality high with about 12.5 percent of the usual point count, beating rival compact methods on quality and memory.

Catalog record

CREATORXianben Yang, Tao Wang, Yuxuan Li, Yi Jin, Haibin Ling
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE32442
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.01884

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

GS^2: Graph-based Spatial Distribution Optimization for Compact 3D Gaussian Splatting Xianben Yang, Tao Wang, Yuxuan Li, Yi Jin, Haibin Ling Abstract 3D Gaussian Splatting (3DGS) has demonstrated breakthrough performance in novel view synthesis and real-time rendering. Nevertheless, its practicality is constrained by the high memory cost due to a huge number of Gaussian points. Many pruning-based 3DGS variants have been proposed for memory saving, but often compromise spatial consistency and may lead to rendering artifacts. To address this issue, we propose graph-based spatial distribution optimization for compact 3D Gaussian Splatting (GS^2), which enhances reconstruction quality by optimizing the spatial distribution of Gaussian points. Specifically, we introduce an evidence lower bound (ELBO)-based adaptive densification strategy that automatically controls the densification process. In addition, an opacity-aware progressive pruning strategy is proposed to further reduce memory consumption by dynamically removing low-opacity Gaussian points. Furthermore, we propose a graph-based feature encoding module to adjust the spatial distribution via feature-guided point shifting. Extensive experiments validate that GS^2 achieves a compact Gaussian representation while delivering superior rendering quality. Compared with 3DGS, it achieves higher PSNR with only about 12.5% Gaussian points. Furthermore, it outperforms all compared baselines in both rendering quality and memory efficiency. The source code is publicly available at https://github.com/BJTU-KD3D/GS-2 . 1 Introduction Novel view synthesis is critical in computer vision and computer graphics, and has attracted great attention in many applications, such as augmented reality 4 and autonomous driving 28 . Neural Radiance Fields (NeRF) 23 generates high-fidelity images through volume rendering, demonstrating excellent detail preservation. However, its time-consuming network inference limits its application in practical scenarios. Recently, 3D Gaussian Splatting (3DGS) 13 has attracted intensive attention due to its promising rendering quality and computational efficiency. 3DGS enhances detail representation through densification, enabling more efficient and higher-quality 3D scene reconstruction than NeRF. Render View GS View (a) LightGaussian (b) Ours Figure 1 : Comparison of rendering quality and spatial distribution. We report the PSNR and number of Gaussian points ( N_{\textrm{GS}} , in millions) for each method. While a SOTA solution (LightGaussian 6 ) compromises spatial consistency with visual artifacts, our method produces a more uniform and coherent distribution. While 3DGS effectively densifies sparse Structure from Motion (SfM) points into a dense set of Gaussian points, the process often leads to over-parameterization, resulting in excessive memory overhead and reduced rendering speed. For example, unbounded 360-degree scenes from the Mip-NeRF 360 dataset 2 typically require over 3 million Gaussian points when reconstructed using 3DGS 13 . Many approaches attempt to alleviate this burden by substantially reducing the number of Gaussian points through pruning 18 ; 20 ; 8 ; 6 . However, pruning Gaussian points often disrupts spatial consistency and continuity, resulting in noticeable rendering artifacts, as illustrated in Fig. 1 . These methods ignore the global and local continuity constraints, leading to noticeable degradation in rendering quality, as shown in Table 1 . To address these challenges, we introduce G raph-based S patial distribution optimization for compact 3D G aussian S platting (GS^2). Our method utilizes a graph-based feature encoding network along with global alignment loss and local smoothness loss to optimize the spatial distribution of Gaussian points after pruning. Specifically, GS^2 first optimizes densification to avoid redundant Gaussian points overgrowth and then performs fast pruning of redundant Gaussian points using an opacity-based loss function. After that, it employs a graph-based feature encoding network to extract spatial features from the pruned Gaussian points, guiding the optimization of its spatial distribution. Meanwhile, the incorporation of global alignment loss and local smoothness loss improves spatial consistency and continuity by encouraging Gaussian points with similar features to remain close in Euclidean space. Additionally, other attributes of the Gaussian points ( e.g. , colors and scale) are jointly updated during spatial optimization. Figure 2 : Comparison of rendering quality (PSNR) and number of Gaussian points ( N_{\textrm{GS}} , in millions ) for various methods on the Mip-NeRF360 dataset. For evaluation, we compare the proposed method with the standard 3DGS and state-of-the-art pruning-based variants, on the challenging Mip-NeRF 360 2 and Tanks & Temples 15 datasets. As shown in Fig. 2 and Table 1 , GS^2 achieves higher PSNR than 3DGS using only about 12.5% Gaussian points. Moreover, it surpasses all compared baselines in both rendering quality and memory efficiency. In summary, our contributions include: • We introduce adaptive densification and opacity-aware pruning to remove Gaussian points and reduce memory consumption with the minimal render quality loss; • we propose a graph-based encoder that leverages feature information to guide the spatial redistribution of Gaussian points, thereby enhancing their spatial consistency and continuity; and • we conduct extensive experiments on two challenging benchmark datasets, showing that our method achieves high rendering quality with dramatically reduced number of Gaussian points. Figure 3 : The pipeline of our method. We propose an Adaptive Densification and Pruning (ADP) module that adaptively increases point density in phase 1 and prunes low-opacity points in phase 2. To improve the spatial distribution after pruning, we introduce a Graph-based Spatial Distribution Optimization (GSDO) module, which refines spatial distribution in phase 3 using a lightweight graph-based feature encoder with global alignment and local smoothness losses. 2 Related Work 2.1 Novel View Synthesis Novel view synthesis involves generating images from perspectives different from the input views by leveraging existing multi-view images. NeRF 23 achieves photorealistic novel view synthesis by leveraging neural networks for implicit scene representation. Despite its impressive rendering performance, NeRF suffers from low computational efficiency due to its reliance on multilayer perceptrons (MLPs). Although several works 11 ; 19 ; 38 ; 30 ; 1 ; 16 ; 32 aim to mitigate these limitations, fundamental issues remain. In contrast, 3DGS 13 has gained significant attention for enabling real-time, high-fidelity rendering. Recent extensions further enhance its applicability, including anti-aliasing techniques 39 ; 33 , memory optimization 6 ; 26 ; 25 ; 18 ; 24 ; 21 , replacement of spherical harmonics for modeling high-frequency reflectance 35 , and dynamic scene modeling 22 ; 36 ; 31 ; 34 ; 12 ; 17 ; 10 . Although 3DGS offers the aforementioned advantages, it still suffers from high memory overhead. 2.2 Compact Gaussian Splatting 3DGS initializes the Gaussian points distribution from a sparse point cloud generated by SfM and employs an adaptive density control mechanism to prune, split, or clone Gaussian points during optimization. Nevertheless, this mechanism is susceptible to the quality of the initial point cloud and prone to producing redundant Gaussian points. To reduce memory overhead, codebook-based compression 6 ; 18 ; 25 ; 26 and pruning strategies 6 ; 18 ; 27 ; 41 ; 9 are proposed to quantize Gaussian parameters or discard non-contributing Gaussian points. Anchor-based methods 21 ; 5 improve spatial efficiency by representing Gaussians with anchors and modeling their relationships to reduce redundancy. Among these approaches, pruning-based 3DGS variants have received particular attention for effectively balancing memory reduction and rendering quality. Pruning-based methods can be typically divided into two groups: those by combining masking with scale or opacity 18 ; 20 , and those by hand-crafted importance scores 8 ; 6 . While these methods achieve competitive compression ratios, they ignore the spatial rearrangement of the remaining Gaussian points after pruning.

Read the complete text