SizeGS: Size-aware Compression of 3D Gaussian Splatting via Mixed Integer Programming

TL;DR

SizeGS addresses the large data footprint of 3D Gaussian Splatting by casting size-aware compression as a MINLP and decoupling the optimization into discrete reserve-ratio search and 0-1 ILP for bit-widths. It establishes a linearized size model and a quantization-loss proxy to rapidly approximate quality, enabling fast exploration of hyperparameters to meet a target size. Through CUDA-accelerated quantization and a piecewise finetuning stage, SizeGS delivers state-of-the-art offline compression and competitive results against online methods after finetuning, across multiple 3DGS variants. The approach offers practical benefits for bandwidth-fluctuating scenarios such as volumetric streaming and remote teleoperation by enabling quick, near real-time size-constrained compression.

Abstract

Recent advances in 3D Gaussian Splatting (3DGS) have greatly improved 3D reconstruction. However, its substantial data size poses a significant challenge for transmission and storage. While many compression techniques have been proposed, they fail to efficiently adapt to fluctuating network bandwidth, leading to resource wastage. We address this issue from the perspective of size-aware compression, where we aim to compress 3DGS to a desired size by quickly searching for suitable hyperparameters. Through a measurement study, we identify key hyperparameters that affect the size -- namely, the reserve ratio of Gaussians and bit-width settings for Gaussian attributes. Then, we formulate this hyperparameter optimization problem as a mixed-integer nonlinear programming (MINLP) problem, with the goal of maximizing visual quality while respecting the size budget constraint. To solve the MINLP, we decouple this problem into two parts: discretely sampling the reserve ratio and determining the bit-width settings using integer linear programming (ILP). To solve the ILP more quickly and accurately, we design a quality loss estimator and a calibrated size estimator, as well as implement a CUDA kernel. Extensive experiments on multiple 3DGS variants demonstrate that our method achieves state-of-the-art performance in post-training compression. Furthermore, our method can achieve comparable quality to leading training-required methods after fine-tuning.

Figures (6)

• Figure 1: The reserve ratio and bit-width strongly impact file size and exhibit a near-linear relationship with it.
• Figure 2: Size-aware compression pipeline. We achieve size-aware 3DGS compression via hyperparameter search. As illustrated above, the compression process can be controlled by adjusting the reserve ratio $\tau$ and bit-width setting ${\mathbf{Q}}$. Algo. \ref{['algo:searching']} presents our hyperparameter optimization algorithm, which aims to meet the size target while maximizing quality. To accelerate the bit-width setting solver, we use 0-1 ILP to compute the bit-width for each channel, and then construct a 0-1 ILP per channel to solve for the bit-width of each group. Additionally, we implement parallel group-wise quantization to further speed up the process. Finally, we restore quality via piecewise fine-tuning after pruning, voxelization, and quantization.
• Figure 3: Qualitative results. We present the rendering results (rows 1) along with the corresponding error maps (rows 2) from a randomly selected viewpoint of the bicycle scene.
• Figure 4: Rate-distortion curves for quantitative comparison. Note that our goal is not to improve the marginal performance and defeat the existing compression works. Instead, we aim to design a hyperparameter parameter searching algorithm to compress the 3DGS model into the desired size while maximizing visual quality.
• Figure 5: Bit widths of each channel. We use different colors to represent different kinds of attributes. From left to right are: features ${\mathbf{f}}$, offsets ${\mathbf{O}}$, opacity $o$, scaling ${\mathbf{l}}$, and rotation ${\mathbf{r}}$.