GaussianPOP: Principled Simplification Framework for Compact 3D Gaussian Splatting via Error Quantification

TL;DR

GaussianPOP introduces an analytically derived per-Gaussian visual error metric, ΔSE_k, directly from the 3DGS rendering equation to enable principled pruning. The authors implement a render-once, compute-locally algorithm for efficient, single-pass error quantification and present a three-stage framework—quantification, pruning, and fine-tuning—that works for both on-training and post-training scenarios. They demonstrate that direct error quantification yields superior compactness–fidelity trade-offs compared with state-of-the-art pruning methods, supported by extensive experiments and an iterative re-quantification strategy for stable post-training simplification. The approach reduces model size while preserving high rendering quality, facilitating deployment of compact 3D Gaussian Splatting representations on resource-constrained devices.

Abstract

Existing 3D Gaussian Splatting simplification methods commonly use importance scores, such as blending weights or sensitivity, to identify redundant Gaussians. However, these scores are not driven by visual error metrics, often leading to suboptimal trade-offs between compactness and rendering fidelity. We present GaussianPOP, a principled simplification framework based on analytical Gaussian error quantification. Our key contribution is a novel error criterion, derived directly from the 3DGS rendering equation, that precisely measures each Gaussian's contribution to the rendered image. By introducing a highly efficient algorithm, our framework enables practical error calculation in a single forward pass. The framework is both accurate and flexible, supporting on-training pruning as well as post-training simplification via iterative error re-quantification for improved stability. Experimental results show that our method consistently outperforms existing state-of-the-art pruning methods across both application scenarios, achieving a superior trade-off between model compactness and high rendering quality.

Figures (9)

• Figure 1: We propose GaussianPOP, enabling high-quality and compact view synthesis with superior rendering of details. The figure compares our method against state-of-the-art approaches in two key scenarios. The on-training scenario integrates pruning into the training process from scratch, whereas the post-training scenario applies simplification to a pre-trained model.
• Figure 2: Overview of the GaussianPOP simplification pipeline. We quantify the per-pixel error ($\Delta SE_k$) for each Gaussian by analytically deriving its visual contribution from the $\alpha$-blending equation. This error is calculated efficiently, allowing us to prune low-error Gaussians to create a compact model.
• Figure 3: Distribution of the cumulative square error $\Delta_{SE}$ across all training views for pre-trained 'Bonsai' scene. The majority of Gaussians contribute near-zero error, indicating that most Gaussians are non-essential to the final render.
• Figure 4: Qualitative comparison for 'Garden' scene at 0.25M Gaussians. The density visualization (top row) illustrates our method achieves a more effective distribution for capturing detailed regions compared to GaussianSpa. This superior distribution results in a final render (bottom row) that preserves finer detail and texture, achieving a higher PSNR.
• Figure 5: Quality-compactness trade-off for on-training pruning. The plots compare PSNR against the number of Gaussians for our method and competing approaches on two representative scenes.