Improving Densification in 3D Gaussian Splatting for High-Fidelity Rendering

TL;DR

The paper tackles the densification bottleneck in 3D Gaussian Splatting (3DGS) by introducing three complementary improvements: when to densify via an Edge-Aware Score (EAS), how to densify using a Long-Axis Split (LAS), and how to mitigate overfitting with Recovery-Aware Pruning (RAP), Multi-View Update (MU), and Growth Control (GC). The proposed framework integrates these components to achieve higher rendering fidelity with fewer Gaussians and without additional training or rendering overhead. Empirical results on real-world datasets (Mip-NeRF 360, Tanks and Temples, Deep Blending) demonstrate state-of-the-art quality (PSNR/SSIM) and competitive perceptual metrics (LPIPS) while reducing computational burden and Gaussian budgets. The approach is evaluated through extensive ablations, showing the individual and combined contributions of EAS, LAS, RAP, MU, and GC, and it is designed for easy integration into existing 3DGS pipelines.

Abstract

Although 3D Gaussian Splatting (3DGS) has achieved impressive performance in real-time rendering, its densification strategy often results in suboptimal reconstruction quality. In this work, we present a comprehensive improvement to the densification pipeline of 3DGS from three perspectives: when to densify, how to densify, and how to mitigate overfitting. Specifically, we propose an Edge-Aware Score to effectively select candidate Gaussians for splitting. We further introduce a Long-Axis Split strategy that reduces geometric distortions introduced by clone and split operations. To address overfitting, we design a set of techniques, including Recovery-Aware Pruning, Multi-step Update, and Growth Control. Our method enhances rendering fidelity without introducing additional training or inference overhead, achieving state-of-the-art performance with fewer Gaussians.

Figures (16)

• Figure 1: Compare the geometric differences before and after splitting between split and LAS. The outer dashed line represents the Gaussian shape before splitting.
• Figure 2: Evaluate the drop in PSNR after splitting using different splitting strategies, test scene is bicycle, a smaller drop indicates less geometric error introduced by the splitting.
• Figure 3: Qualitative comparison results among scenes garden, drjohnson, train.
• Figure 4: Qualitative comparison for evaluating the effectiveness of EAS.
• Figure 5: Evaluating the impact of LAS on optimization speed.