T L D R :
We model the 3D space with the abstract Gaussian splatting, which provides compact and efficient representations with high fidelity.
We model the 3D space with the abstract Gaussian splatting, which provides compact and efficient representations with high fidelity.
We present Smol-GS, a novel method for learning compact representations for 3D Gaussian Splatting (3DGS). Our approach learns highly efficient encodings in 3D space that integrate both spatial and semantic information. The model captures the coordinates of the splats through a recursive voxel hierarchy, while splat-wise features store abstracted cues, including color, opacity, transformation, and material properties. This design allows the model to compress 3D scenes by orders of magnitude without loss of flexibility. Smol-GS achieves state-of-the-art compression on standard benchmarks while maintaining high rendering quality. Beyond visual fidelity, the discrete representations could potentially serve as a foundation for downstream tasks such as navigation, planning, and broader 3D scene understanding.
Method overview of Smol-GS: We train (trained parameters in blue) neural splats with tiny MLP-decoders for view-dependent rendering. The coordinates are compressed with occupancy-octree coordinate coding. We also learn the quantization and arithmetic coding of splat features with NLL rate terms, and employ adaptive density control with stage-wise training (loss-terms in red) to balance fidelity and size.
These charts for the MipNeRF-360 data set show how different 3DGS compression methods trade model size for reconstruction quality (PSNR, SSIM, LPIPS). Hover over a point to see details, and click legend entries to toggle methods. Smol-GS provides an excellent trade-off in storage size and quality. Smol-GS (tiny--small--base) consistently lies on or beyond the Pareto frontier, achieving markedly smaller sizes at comparable quality to strong baselines collected via the 3DGS.zip benchmark.
These videos show side-by-side novel-view sweeps for several Mip-NeRF 360 scenes, comparing a vanilla 3DGS model trained for 30k steps against our compressed Smol-GS representation. Smol-GS keeps the underlying Gaussian primitive structure but replaces heavy per-splat SH parameters with low-dimensional abstract features decoded by tiny MLPs and stores geometry in an entropy-coded occupancy octree. Despite being about two orders of magnitude smaller than the original 3DGS models, Smol-GS faithfully preserves fine geometry and view-dependent effectsâspecular highlights, glass, and low-texture wallsâand often removes artefacts visible in 3DGS-30K. Together with the quantitative plots below, these sweeps highlight that explicit-but-packed geometry plus learned abstract splat features yield a very strong rateâdistortion trade-off while remaining compatible with downstream 3DGS-style editing.
Qualitative comparison: We visualize the reconstruction result of 3D Gaussian Splatting (3DGS-30K), Smol-GS-base, and the ground truth (GT).
Feature space: We visualize the reconstruction results of Smol-GS-base in the RGB and feature spaces. The 8-dim features are mapped to RGB color by RGB=sigmoid(PCA(feat)).
