PCGS: Progressive Compression of 3D Gaussian Splatting

TL;DR

PCGS tackles the high data size of 3D Gaussian Splatting by introducing progressive compression that jointly tunes anchor quantity and quality. It employs a rate-aware masking strategy to progressively add anchors and refine Gaussians, and a progressive quantization scheme (Round followed by trit-plane quantization) with a context-aware, trinomial entropy model to improve efficiency across levels. The framework unifies training into a single process that yields a full rate-distortion curve and enables on-demand bitstream refinement without retraining. Experiments across multiple large-scale datasets demonstrate that PCGS achieves comparable compression to state-of-the-art single-rate methods while providing scalable, progressively refinable bitstreams suitable for dynamic bandwidth and storage constraints. This enhances the practical applicability of 3DGS in real-world pipelines and streaming scenarios.

Abstract

3D Gaussian Splatting (3DGS) achieves impressive rendering fidelity and speed for novel view synthesis. However, its substantial data size poses a significant challenge for practical applications. While many compression techniques have been proposed, they fail to efficiently utilize existing bitstreams in on-demand applications due to their lack of progressivity, leading to a waste of resource. To address this issue, we propose PCGS (Progressive Compression of 3D Gaussian Splatting), which adaptively controls both the quantity and quality of Gaussians (or anchors) to enable effective progressivity for on-demand applications. Specifically, for quantity, we introduce a progressive masking strategy that incrementally incorporates new anchors while refining existing ones to enhance fidelity. For quality, we propose a progressive quantization approach that gradually reduces quantization step sizes to achieve finer modeling of Gaussian attributes. Furthermore, to compact the incremental bitstreams, we leverage existing quantization results to refine probability prediction, improving entropy coding efficiency across progressive levels. Overall, PCGS achieves progressivity while maintaining compression performance comparable to SoTA non-progressive methods. Code available at: github.com/YihangChen-ee/PCGS.

Figures (7)

• Figure 1: Comparison of existing approaches (upper) and the proposed progressive compression (lower). Existing approaches generate multiple independent bitstreams targeting different rates and fidelity through multiple trainings, while a progressive compression approach (with only one training) can continuously improve the fidelity by incrementally adding bitstreams, which is resource-saving in on-demand applications.
• Figure 2: Overview of the proposed PCGS, which controls anchors in both quantity and quality in a progressive way, i.e., progressively decoding new anchors via masking control and refining existing anchors with finer quantization steps. Left: Given anchor $\bm{x}^\text{a}$, its position is interpolated within the binary hash grid to obtain the hash feature $\bm{f}^\text{h}$. A rate-aware MLP, conditioned on the level information $\lambda_s$, utilizes $\bm{f}^\text{h}$ to determine the quantization steps and provide context information for different progressivity levels. Additionally, the anchor and Gaussian masks $\bm{m}^a_{s}$ and $\bm{m}^g_{s}$ are obtained via the rate-aware progressive masking strategy from learnable features $\bm{f}^\text{m}_{s}$ and $\bm{f}^\text{m}_{base}$ (bottom). Right: At each level $s$, according to its mask $\bm{m}[i]^\text{a}_s$, the $i$-th anchor either remains undecoded when $\bm{m}[i]^\text{a}_s=0$, be newly decoded when $\bm{m}[i]^\text{a}_s$ transits from $0$ to $1$, or otherwise be refined.
• Figure 3: Progressive quantization. (a): Illustration of progressive quantization. The real value $\bm{f}$ is initially quantized at level $s=1$ using $\texttt{Round}$ with the step size $q_{1}$. It is then progressively refined using trit-plane quantization at levels $s=2$, and $3$, gradually approaching $\bm{f}$. (b): Three possible paths to reach the quantized value $\hat{\bm{f}}_3$ at the current level ($s=3$) in training with $\Delta\bm{m}^\text{a}_2=1$ indicating that the middle path is selected by mask weighting.
• Figure 4: R-D curve comparison of different methods. The number of training iterations for each method is indicated in parentheses. For GoDe, it applies an additional finetune stage of $30k$ iterations to the Scaffold-GS which is originally trained for $30k$. Diamond markers $\Diamond$ represent progressive compression methods, while circle markers $\bigcirc$ denote traditional single-rate compression methods. More results can be obtained in the Appendix.
• Figure 5: R-D curve results of ablation study. Experiments are conducted on Mip-NeRF360 mip360.