Smol-GS Compresses 3D Gaussian Splatting Scenes by Orders of Magnitude Without Quality Loss

Researchers from Aalto University and ELLIS Institute Finland have introduced Smol-GS, a compression method for 3D Gaussian Splatting (3DGS) that dramatically reduces model size while preserving rendering fidelity. The approach encodes scene geometry using a recursive occupancy-octree coordinate hierarchy and replaces per-splat spherical harmonic parameters with low-dimensional abstract features decoded by lightweight MLPs — achieving compression of roughly two orders of magnitude compared to standard 3DGS models.

On the MipNeRF-360 benchmark, Smol-GS sits at or beyond the Pareto frontier across PSNR, SSIM, and LPIPS metrics relative to existing compression baselines. Beyond visual fidelity, the authors suggest the discrete, compact representations could support downstream tasks such as navigation and 3D scene understanding, and remain compatible with standard 3DGS-style editing workflows.