GS^2: Graph-based Spatial Distribution Optimization for Compact 3D Gaussian Splatting
Authors: Xianben Yang, Tao Wang, Yuxuan Li, Yi Jin, Haibin Ling
Organizations: Key Laboratory of Big Data and Artificial Intelligence in Transportation, Ministry of Education; School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044, China · Department of Artificial Intelligence, Westlake University, Hangzhou 310030, China
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\textasciicircum2), 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\textasciicircum2 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.
Figures & tables
Figure 1: Comparison of rendering quality and spatial distribution. We report the PSNR and number of Gaussian points ( NGS , 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.
Figure 2: Comparison of rendering quality (PSNR) and number of Gaussian points ( NGS , in millions ) for various methods on the Mip-NeRF360 dataset.
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.
Figure 4: ELBO-guided adaptive densification.
Figure 5: Analyzing the impact of optimization. Low-opacity Gaussian points ( opacity≤0.1 ) constitute a considerable portion of the total. For instance, in the Lighthouse scene from the Tanks & Temples dataset, they account for approximately 40% of all Gaussian points.
Figure 6: Impact of pruning. We report PSNR and NGS (in millions) for each method on the church scene from Tanks & Temples. Pruning degrades the spatial consistency and continuity of the Gaussian points, leading to visible artifacts.
Dataset
Method
All Scenes
Outdoor
Indoor
PSNR ↑
SSIM ↑
LPIPS ↓
NGS↓
PSNR ↑
SSIM ↑
LPIPS ↓
NGS↓
PSNR ↑
SSIM ↑
LPIPS ↓
NGS↓
3DGS
27.71
0.83
0.20
3.12
24.99
0.74
0.23
4.50
31.09
0.92
0.19
1.41
Mip-NeRF 360
EAGLES
27.22
0.82
0.23
1.25
24.64
0.73
0.27
1.69
30.46
0.93
0.18
0.71
Mini-Splatting
27.34
0.82
0.22
0.49
24.71
0.74
0.24
0.57
30.61
0.92
0.19
0.40
CompGS
27.25
0.82
0.23
0.86
24.79
0.73
0.26
1.19
30.33
0.92
0.19
0.46
Compact 3DGS
27.25
0.81
0.23
1.35
24.61
0.72
0.27
1.47
30.54
0.92
0.18
1.21
Table 1: Quantitative rendering results of Mip-NeRF 360 and T&T. Bold font highlights the best results and underline for the second.
Figure 7: Qualitative comparison. The first scene ( Room ) is from the Mip-NeRF 360 dataset, followed by four scenes ( Palace, Temple, Lighthouse, Courtroom ) from the Tanks & Temples dataset. Key regions have been zoomed in to highlight visual differences. We report the PSNR and NGS (in millions) for each method.
Dataset
Method
PSNR ↑
SSIM ↑
LPIPS ↓
NGS↓
Mip-NeRF 360
3DGS
27.71
0.83
0.20
3.12
+ ELBO densification
27.73
0.83
0.20
2.82
+ Opacity pruning
26.79
0.80
0.26
0.30
+ Increase iterations
26.98
0.80
0.26
0.30
+ Lsmt
27.53
0.81
0.25
0.30
Full Model
27.74
0.82
0.23
0.30
Table 2: Ablation study. Evaluate the effectiveness of each component of our model on rendering quality.
Dataset
Method
PSNR ↑
SSIM ↑
LPIPS ↓
NGS↓
Mip-NeRF 360
LightGaussian
27.35
0.82
0.23
1.07
(+GSDO)
27.75
0.82
0.21
1.07
MaskGaussian
27.61
0.82
0.23
1.50
(+GSDO)
27.94
0.83
0.20
1.50
T&T
LightGaussian
23.16
0.82
0.25
0.53
(+GSDO)
23.97
0.83
0.22
0.53
Table 3: Integration of GSDO with other 3DGS variants.
Dataset
Method
Train Time (min) ↓
FPS ↑
T&T
3DGS
16
135
EAGLES
21
247
Mini-Splatting
14
576
CompGS
17
325
Compact 3DGS
27
345
LightGaussian
17
214
Table 4: Training time and FPS comparison on T & T.
Dataset
Interval
PSNR ↑
SSIM ↑
LPIPS ↓
Mip-NeRF 360
100
27.74
0.82
0.23
200
27.65
0.82
0.24
400
27.60
0.82
0.24
Table 5: Ablation study for different pruning intervals.
Recent significant advances in 3D scene representation have been driven by 3D Gaussian Splatting (3DGS), which has enabled real-time rendering with photorealistic quality. 3DGS often requires a large number of primitives to achieve high fidelity, leading to redundant representations and high resource consumption, thereby limiting its scalability for complex or large-scale scenes. Consequently, effective pruning strategies and more expressive primitives that can reduce redundancy while preserving visual quality are crucial for practical deployment. We propose an efficient, integrated reconstruction-aware pruning strategy that adaptively determines pruning timing and refining intervals based on reconstruction quality, thus reducing model size while enhancing rendering quality. Moreover, we introduce a 3D Difference-of-Gaussians primitive that jointly models both positive and negative densities in a single primitive, improving the expressiveness of Gaussians under compact configurations. Our method significantly improves model compactness, achieving up to 90% reduction in Gaussian-count while delivering visual quality that is similar to, or in some cases better than, that produced by state-of-the-art methods. Code will be made publicly available.
Haoran Wang, Guoxi Huang, Fan Zhang +2
School of Computer Science, University of Bristol, Bristol, UK
While 3D Gaussian Splatting (3DGS) has revolutionized 3D reconstruction, it suffers from significant overhead due to massive redundant primitives. Existing compression methods typically rely on local sampling or fixed pruning thresholds, which often struggle to balance redundancy reduction with high-fidelity rendering. To address this, we propose a novel framework that formulates Gaussian optimization as a global geometric distribution matching problem. Specifically, our approach integrates three components: (1) we introduce a multi-view 3D Gaussian contribution ranking mechanism that filters primitives using geometric consistency instead of local heuristics; (2) we propose a global Optimal Transport (OT)-based aggregation algorithm that merges redundant primitives while preserving the underlying geometry; and (3) we design an OT-based densification operator that maintains the Gaussian's distributional properties for stable optimization. Our approach achieves state-of-the-art rendering quality with only \textbf{10%} primitives and \textbf{10×} accelerated training speeds compared to vanilla 3DGS.
Beizhen Zhao, Sicheng Yu, Ziran Yin +2
The Hong Kong University of Science and Technology (Guangzhou)
3D Gaussian Splatting (3DGS) enables real-time novel view synthesis, but existing general-purpose acceleration methods suffer severe rendering quality degradation when extended to more complex, large-scale scenes. To address this issue, we propose EffGS, a more general acceleration framework that improves training and rendering efficiency while maintaining reconstruction quality comparable to or better than vanilla 3DGS across bounded and large-scale scenes. EffGS combines frequency-aware guidance, localized density control, and adaptive primitive scale modulation. First, an importance scoring mechanism combines pixel-wise reconstruction errors with a difference-of-Gaussians mask scheduled over training to provide stage-dependent spatial guidance. Second, localized densification and pruning restricts density modifications to Gaussians with valid projected footprints in the sampled views. Third, learnable per-Gaussian scale modulation adjusts effective primitive extent during optimization while retaining the Compact Box rasterization rule. Extensive experiments on bounded and large-scale scene datasets demonstrate a favorable balance between reconstruction quality, training time, and primitive count. Component ablations and matched-primitive-budget comparisons further support the effectiveness of the framework.
Changbai Li, Shuo Yang, Yichen Yang +2
Beihang University · Nanyang Technological University