Abstract

3D Gaussian Splatting (3DGS) enables real-time novel-view synthesis through explicit Gaussian representations and differentiable rasterization, yet its model size—often hundreds of megabytes per scene—severely limits deployment on mobile devices, web applications, and AR/VR platforms. This paper presents a systematic compression study of 3DGS on the NeRF Synthetic Lego scene using only an Apple M4 consumer laptop with the Metal Performance Shaders (MPS) backend. We first train a lightweight baseline (1,000 Gaussians, spherical-harmonics degree 1, 200×200 rendering resolution), and then evaluate six compression configurations based on spherical-harmonics (SH) distillation, opacity pruning, and vector quantization. All results are measured on 200 test views using real PSNR, SSIM, and LPIPS scores. Our experiments show that reducing the SH degree from 1 to 0 yields 1.64× compression with virtually no quality loss. Combining SH degree 0 with opacity pruning at threshold τ = 0.02 achieves a 9.08× compression ratio, reducing the model from 0.0877 MB to 0.0097 MB (88.9% size reduction). These findings validate the effectiveness of simple post-hoc compression strategies in resource-constrained settings and provide practical guidance for choosing quality–size trade-offs.

Keywords