Sort-free Gaussian Splatting via Weighted Sum Rendering

TL;DR

This work addresses the sorting bottleneck in 3D Gaussian Splatting by introducing Sort-free Gaussian Splatting via Weighted Sum Rendering (GS-WSR), which replaces non-commutative alpha blending with commutative weighted sums. It presents three WSR variants (DIR-WSR, EXP-WSR, LC-WSR) and a view-dependent opacity mechanism based on learnable spherical harmonics, enabling high-quality rendering without depth ordering. The method is trained end-to-end in a differentiable framework and demonstrated on mobile GPUs with Vulkan/GLSL, achieving about 1.23× faster rendering on average while maintaining competitive image quality against state-of-the-art methods. The key contributions are the first sort-free Gaussian Splatting approach compatible with graphics pipelines, the Weighted Sum Rendering formulation, and the demonstrated practical efficiency gains on resource-constrained devices.

Abstract

Recently, 3D Gaussian Splatting (3DGS) has emerged as a significant advancement in 3D scene reconstruction, attracting considerable attention due to its ability to recover high-fidelity details while maintaining low complexity. Despite the promising results achieved by 3DGS, its rendering performance is constrained by its dependence on costly non-commutative alpha-blending operations. These operations mandate complex view dependent sorting operations that introduce computational overhead, especially on the resource-constrained platforms such as mobile phones. In this paper, we propose Weighted Sum Rendering, which approximates alpha blending with weighted sums, thereby removing the need for sorting. This simplifies implementation, delivers superior performance, and eliminates the "popping" artifacts caused by sorting. Experimental results show that optimizing a generalized Gaussian splatting formulation to the new differentiable rendering yields competitive image quality. The method was implemented and tested in a mobile device GPU, achieving on average $1.23\times$ faster rendering.

Figures (7)

• Figure 1: The architecture of sort-free Gaussian Splatting via Weighted Sum Rendering. 3DGS needs to tiling, replication, sorting, and rendering. Our method only needs to calculate the weight for each Gaussian, and independently sum their contributions per-pixel.
• Figure 2: Three variants of Weighted Sum Rendering with different weight calculations, namely Direct Weighted Sum Rendering (DIR-WSR), Exponential Weighted Sum Rendering (EXP-WSR), and Linear Correction Weighted Sum Rendering (LC-WSR).
• Figure 3: View dependent opacity. In 3DGS, the accumulated $\hat{\alpha}$ changes depending on the viewer's direction, which motivates us to assign view-dependent opacities in sort-free Gaussian Splatting.
• Figure 4: Visual comparison on the Mip-NeRF 360 dataset. Our OIT method achieves similar visual performance compared to 3DGS. Please note that our method doesn’t require the order of Gaussians.
• Figure 5: Our method eliminates the "popping" artifacts during view transformation.