Gradient-Driven Natural Selection for Compact 3D Gaussian Splatting

TL;DR

This work tackles the storage and compute burden of 3D Gaussian Splatting (3DGS) by introducing a natural-selection–inspired pruning framework. A globally applied regularization gradient on opacities competes with rendering- quality gradients to autonomously determine which Gaussians to retain, eliminating the need for manual pruning rules or extra parameters. A finite-prior opacity decay accelerates pruning without sacrificing quality, enabling state-of-the-art performance with only 15% of the original Gaussian budget and faster convergence. The approach yields more uniform Gaussian distributions, preserves rendering fidelity, and is readily portable to advanced 3DGS variants, offering practical improvements for compact, real-time 3D scene representations.

Abstract

3DGS employs a large number of Gaussian primitives to fit scenes, resulting in substantial storage and computational overhead. Existing pruning methods rely on manually designed criteria or introduce additional learnable parameters, yielding suboptimal results. To address this, we propose an natural selection inspired pruning framework that models survival pressure as a regularization gradient field applied to opacity, allowing the optimization gradients--driven by the goal of maximizing rendering quality--to autonomously determine which Gaussians to retain or prune. This process is fully learnable and requires no human intervention. We further introduce an opacity decay technique with a finite opacity prior, which accelerates the selection process without compromising pruning effectiveness. Compared to 3DGS, our method achieves over 0.6 dB PSNR gain under 15\% budgets, establishing state-of-the-art performance for compact 3DGS. Project page https://xiaobin2001.github.io/GNS-web.

Figures (12)

• Figure 1: Our method's pipeline consists of two parts. The upper section shows the full workflow: starting from sparse SfM point clouds, we densify and optimize to obtain a high-quality dense scene, which is then refined via a natural selection mechanism to produce a compact, high-fidelity representation. The lower section details this natural selection framework: a globally consistent regularization gradient field (dashed lines) is applied to all Gaussian opacities, guiding optimization gradients (solid lines) to identify and prune Gaussians whose opacity falls below a survival threshold. Those Gaussians with smaller optimized gradients will gradually decay in opacity until reaching the death threshold, after which they are permanently removed. The result is a high-quality, compact scene.
• Figure 2: Qualitative comparison between our method and 3DGS on the garden scene
• Figure 3: The figure illustrating the optimization speed improvement of finite prior over no prior.
• Figure 4: Performance under varying Gaussian budgets. Results for the remaining scenes can be found in appendix.
• Figure 5: Qualitative comparison results among scenes garden, room, train. Num is final Gaussian count after training. The results for the remaining scenes can be found in appendix.