Prune Wisely, Reconstruct Sharply: Compact 3D Gaussian Splatting via Adaptive Pruning and Difference-of-Gaussian Primitives

TL;DR

This work proposes 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 and introduces 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.

Abstract

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.

Figures (7)

• Figure 1: Performance improvement of the proposed pruning strategy and 3D Difference-of-Gaussian (DoG) primitives. Comparison on the Drjohnson scene, where the original 3DGS method is shown on the left and our result in the middle panel. The right panel shows PSNR versus the number of primitives on the Tanks and Temples dataset knapitsch2017tanks across various 3DGS-based methods.
• Figure 2: (Left) Gaussian primitive count comparison. Our method adaptively adjusts the refinement settings to meet different pruning targets, such as the 50% and 90% pruning ratios shown in the figure. (Right) Overview of the Reconstruction-aware Pruning Scheduler and 3D-DoG Density Control. We use L1 loss as a reconstruction quality indicator to dynamically determine pruning timing and ratio throughout optimization. In addition, we activate 3D-DoG after pruning and adaptively control its density.
• Figure 3: (Top) Illustration of the proposed 3D-DoG primitive in 1D and 3D, featuring a positive-density peak and a negative-density ring. (Bottom) 3DGS with 3D-DoG primitives achieves better detail representation.
• Figure 4: Novel view rendering comparison with the baselines. Top: Train from the Tanks & Temples. Middle: Playroom from the Deep Blending dataset. Bottom: Treehill from the Mip-NeRF 360 dataset. We have shown details below the images. Best viewed when zoomed in.
• Figure 5: Variations in the primitive count and PSNR values of (Left) Bicycle and (Right) Room scenes using our method. The PSNR drop at the 25k iteration is due to the activation of 3D-DoG.