ControlGS: Consistent Structural Compression Control for Deployment-Aware Gaussian Splatting

TL;DR

ControlGS tackles deployment challenges in 3D Gaussian splatting by providing a scene-agnostic control mechanism that balances rendering fidelity and model size. It achieves this with Uniform Splitting and Opacity-based Sparsification unified under a single control hyperparameter, driving optimization toward an efficient regime across diverse scenes. Empirical results across multiple datasets show that ControlGS delivers higher fidelity with fewer Gaussians than baselines and provides smooth, cross-scene adjustable trade-offs suitable for automated deployment. This work advances practical real-time NVS by enabling consistent, deployment-aware structural compression.

Abstract

3D Gaussian Splatting (3DGS) is a highly deployable real-time method for novel view synthesis. In practice, it requires a universal, consistent control mechanism that adjusts the trade-off between rendering quality and model compression without scene-specific tuning, enabling automated deployment across different device performances and communication bandwidths. In this work, we present ControlGS, a control-oriented optimization framework that maps the trade-off between Gaussian count and rendering quality to a continuous, scene-agnostic, and highly responsive control axis. Extensive experiments across a wide range of scene scales and types (from small objects to large outdoor scenes) demonstrate that, by adjusting a globally unified control hyperparameter, ControlGS can flexibly generate models biased toward either structural compactness or high fidelity, regardless of the specific scene scale or complexity, while achieving markedly higher rendering quality with the same or fewer Gaussians compared to potential competing methods. Project page: https://zhang-fengdi.github.io/ControlGS/

Figures (9)

• Figure 1: Gaussian count--rendering quality relationship across scenes. Left six panels: Empirical curves obtained using the top-performing budget-based method 3DGS-MCMC 3DGS-MCMC, covering representative object, indoor, and outdoor scenes from multiple benchmark datasets GigaNVSMip-NeRF360TanksAndTemplesNeRF. Although the absolute number of Gaussians varies across scenes due to differences in scale, the resulting relationships consistently exhibit a universal four-phase pattern. Right: Conceptual illustration of the four-phase pattern: (A) underfitting, (B) efficient regime, (C) saturation, and (D) overfitting.
• Figure 2: Overview of the ControlGS pipeline. Training starts from an SfM-initialized Gaussian set and proceeds with RGB reconstruction and opacity-based sparsification, which prune low-contribution Gaussians and compact the representation. When pruning saturates, indicating convergence at the current resolution, uniform splitting expands candidates and refines structure, after which optimization resumes and the cycle repeats. A smaller $\lambda_\alpha$ retains more candidates for higher rendering quality, whereas a larger $\lambda_\alpha$ enforces stronger sparsification with fewer Gaussians, enabling consistent structural compression control across scenes.
• Figure 3: Cross-scene structural compression control results. Center: Average PSNR of our method versus the control hyperparameter $\lambda_\alpha$; bubble size denotes the average number of Gaussians. As $\lambda_\alpha$ increases, the model becomes more compact, while smaller $\lambda_\alpha$ preserves more Gaussians for higher fidelity. By tuning $\lambda_\alpha$, our method enables smooth and predictable preference control between model compactness and rendering quality. Surrounding plots: Average PSNR versus average Gaussian count on NeRF Synthetic NeRF (object), Mip-NeRF360 Mip-NeRF360 (indoor/outdoor), Tanks & Temples TanksAndTemples (outdoor), Deep Blending DeepBlending (indoor), and GigaNVS GigaNVS (large outdoor), comparing our method (red), 3DGS-MCMC 3DGS-MCMC (blue), which is used to obtain the Gaussian count--rendering quality curve for reference, and LP-3DGS LP-3DGS (gray). All methods are trained for 100k iterations. For ours, $\lambda_\alpha$ ranges from $1\text{e-7}$ to $1\text{e-6}$ with a step of $1\text{e-7}$ (10 control points); LP-3DGS varies its pruning ratio from 0.1 to 0.9 in 0.1 steps (9 control points). 3DGS-MCMC fits the overall Gaussian count--quality curve with scene-dependent budgets (50k--100k for NeRF Synthetic, 100k--700k for others). ControlGS consistently reaches the efficient regime and delivering comparable or higher PSNR with fewer Gaussians than LP-3DGS, while LP-3DGS saturates early and 3DGS-MCMC requires scene-specific tuning.
• Figure 4: Gaussian count--rendering quality (PSNR) relationships across representative scenes of different scales from multiple benchmark datasets GigaNVSMip-NeRF360TanksAndTemplesNeRF. Results are shown for our method (red), the budget-based method 3DGS-MCMC 3DGS-MCMC (blue), and the non-budget-based method LP-3DGS LP-3DGS (gray). The experimental settings follow the same configuration as in Fig. \ref{['fig:count_quality_control']}.
• Figure 5: Opacity distributions of our method (red) and vanilla 3DGS (gray) on Treehill and Flowers scenes from Mip-NeRF360 Mip-NeRF360, with highlighted peaks and opacities.