Modern methods of reconstructing scenes with high fidelity use 3D Gaussian splat representations, enabling novel view synthesis from sparse image sets. However, the memory requirements and rendering efficiency of these representations pose challenges for applications like network streaming and rendering on low-power devices.

To address these challenges, the authors propose a compressed 3D Gaussian splat representation method. This method utilizes sensitivity-aware vector clustering combined with quantization-aware training to compress directional colors and Gaussian parameters efficiently. Through this approach, the learned codebooks achieve a compression rate of up to 31 times on real-world scenes with minimal degradation in visual quality.

The compressed splat representation is demonstrated to be efficiently rendered using hardware rasterization on lightweight GPUs, achieving up to four times higher frame rates than reported via an optimized GPU compute pipeline. The proposed method undergoes extensive experimentation across multiple datasets, demonstrating its robustness and rendering speed.

In terms of compression, the authors tackle the storage and rendering issues of 3D Gaussian Splatting (3DGS) by compressing the reconstructed scene parameters and rendering the compressed representation through GPU rasterization. This compression pipeline involves sensitivity-aware clustering, quantization-aware fine-tuning, and entropy encoding, effectively reducing memory consumption while preserving essential scene details.

Rendering of Gaussians is achieved via GPU rasterization, where each Gaussian is rendered in sorted order as a planar quad. A vertex shader computes screen space vertex positions, while a pixel shader handles color computation and blending, ensuring efficient rendering with high visual fidelity. Implemented with the WebGPU graphics API in Rust, this renderer can run in modern web browsers on a wide range of devices, making it accessible for various applications.