Hybrid Pruning for 3D Gaussian Splatting

Tarun Gangadhar Vadaparthi
đŸ“„ Write up (PDF) đŸ’» Code

Demo Video

Abstract

We propose a simple hybrid pruning strategy for 3D Gaussian Splatting (3DGS). Each splat is ranked by a fused signal comprising SH energy, multi-view projected area, a visibility proxy, and local density after robust normalization. Keeping a budgeted top-K and applying a light cleanup removes floaters while preserving thin structures. Compared to single-signal baselines (opacity/area/energy), the method yields improved rate–distortion while remaining easy to reproduce.

Method Overview

Method Overview Diagram
Overview of our pruning pipeline. Each Gaussian splat is evaluated across multiple signals, fused via z-score normalization and weighted ranking. Top-k splats are retained, followed by refinement to remove floaters, yielding compact yet faithful 3DGS scenes.

Visual Comparison

Original Pruned
Original
Pruned

Results

Rate–distortion curve (PSNR vs kept fraction)
Qualitative comparison grid