Reducing the Memory Footprint of 3D Gaussian Splatting

Analysis

Why This Paper Matters

3D Gaussian splatting has emerged as a leading method for novel view synthesis, offering state-of-the-art visual quality with fast training and real-time rendering. However, its practical adoption is severely limited by enormous memory requirements—scenes often require gigabytes of storage, making transmission and mobile deployment impractical. This paper directly addresses that bottleneck, achieving a 27x compression without sacrificing quality, which is a critical step toward democratizing high-quality 3D rendering on consumer devices.

The significance is amplified by the method's compatibility with existing 3D Gaussian splatting pipelines. Rather than proposing a fundamentally new representation, the authors introduce three lightweight, plug-and-play techniques that can be applied post-training. This pragmatic approach means the work can be immediately adopted by practitioners, accelerating the path from research to real-world applications like AR/VR, mobile gaming, and remote visualization.

Technical Contributions

• Resolution-aware primitive pruning: The authors observe that many Gaussian primitives are redundant at the target rendering resolution. They propose a simple yet effective pruning criterion based on each primitive's projected size relative to the output resolution, removing approximately half of the primitives with negligible visual impact.
• Adaptive spherical harmonic coefficients: Instead of using a fixed number of spherical harmonic (SH) coefficients for all primitives, the method adaptively selects the number of coefficients per primitive based on its contribution to directional radiance. This reduces storage for low-frequency primitives while preserving high-frequency details where needed.
• Codebook-based quantization: The authors apply vector quantization to Gaussian attributes (position, scale, rotation, opacity, SH coefficients) using a learned codebook, combined with half-float (16-bit) representation for remaining attributes. This achieves significant compression beyond standard float32 storage.

Results

On standard datasets (e.g., Mip-NeRF 360, Tanks and Temples), the combined method achieves a 27x reduction in disk size compared to the original 3D Gaussian splatting. Rendering speed improves by 1.7x due to the reduced primitive count. Visual quality metrics (PSNR, SSIM, LPIPS) remain nearly identical to the uncompressed baseline. The authors also demonstrate practical benefits: download times on a mobile device drop from minutes to seconds, enabling real-time streaming of high-quality 3D scenes.

Significance

This work directly addresses the most critical practical limitation of 3D Gaussian splatting—its memory footprint. By enabling efficient storage and transmission, it opens the door to widespread deployment in bandwidth-constrained environments like mobile AR/VR, web-based 3D viewers, and cloud-rendering pipelines. The techniques are general and could inspire similar compression strategies for other point-based or volumetric rendering methods. Moreover, the 1.7x rendering speedup is a welcome side benefit, making real-time applications even more performant. As AI-driven 3D content creation matures, memory-efficient representations like this will be essential for scaling to consumer-grade hardware.