Introduction
3D Gaussian Splatting (3DGS) has revolutionized novel view synthesis by combining high visual quality with real-time rendering speeds. However, 3DGS faces a significant scalability challenge: a single complex scene can require millions of Gaussians, resulting in model sizes reaching several gigabytes. This makes them impractical for mobile devices, web streaming, or large-scale storage. Smol-GS addresses this "weight" problem by introducing a novel compression framework that slashes the memory footprint of 3DGS by over 150x while maintaining high-fidelity rendering.
Methodology
The Smol-GS architecture achieves extreme compactness through four key components:
- Occupancy-Octree Coordinate Coding: Instead of storing raw 32-bit floating-point coordinates, Smol-GS uses a recursive voxel hierarchy. This bitstream is Huffman-coded, providing nearly lossless spatial storage at a fraction of the original size.
- Abstract Splat Features: Each splat carries a low-dimensional (typically 8D) "abstract feature" vector. This acts as a compressed latent code rather than representing color or shape directly.
- Tiny MLP Decoders: Four very small Multi-Layer Perceptrons (MLPs) act as interpreters, taking the abstract feature and viewing direction to reconstruct rendering parameters such as opacity, color, orientation, and scale.
- Learned Quantization and Entropy Coding: Abstract features are quantized using learned step sizes and compressed via arithmetic coding to maximize information density.
Key Results
Smol-GS was evaluated on standard benchmarks including Mip-NeRF 360, Tanks and Temples, and Deep Blending. The results demonstrate that the framework achieves state-of-the-art compactness without sacrificing performance:
- Extreme Compression: On the Mip-NeRF 360 benchmark, Smol-GS compressed a vanilla 3DGS model from 734 MB down to 4.75 MB while maintaining comparable Peak Signal-to-Noise Ratio (PSNR).
- High Performance: Despite the use of MLP decoders, the model maintains impressive rendering speeds, averaging 245.7 FPS on an NVIDIA H200 GPU.
- Visual Fidelity: Qualitatively, Smol-GS preserves sharp edges and complex material effects, such as specular highlights, better than many larger baselines.
Conclusion and Future Work
Smol-GS demonstrates that geometry is highly compressible and that explicit spatial structures, when paired with octrees and MLP-based abstraction, lead to superior compression ratios. Future directions include integrating these "abstract features" as 3D information tokens for Large Language Models (LLMs) and utilizing depth estimation to improve the initial distribution of splats. While the multi-stage training pipeline currently takes approximately 32 minutes per scene, the resulting efficiency opens new doors for 3D content on resource-constrained devices.
Summary updated