Consistent Quantity-Quality Control across Scenes for Deployment-Aware Gaussian Splatting (2505.10473v2)

Abstract: To reduce storage and computational costs, 3D Gaussian splatting (3DGS) seeks to minimize the number of Gaussians used while preserving high rendering quality, introducing an inherent trade-off between Gaussian quantity and rendering quality. Existing methods strive for better quantity-quality performance, but lack the ability for users to intuitively adjust this trade-off to suit practical needs such as model deployment under diverse hardware and communication constraints. Here, we present ControlGS, a 3DGS optimization method that achieves semantically meaningful and cross-scene consistent quantity-quality control. Through a single training run using a fixed setup and a user-specified hyperparameter reflecting quantity-quality preference, ControlGS can automatically find desirable quantity-quality trade-off points across diverse scenes, from compact objects to large outdoor scenes. It also outperforms baselines by achieving higher rendering quality with fewer Gaussians, and supports a broad adjustment range with stepless control over the trade-off. Project page: https://zhang-fengdi.github.io/ControlGS/

• The paper introduces ControlGS, a novel 3D Gaussian Splatting method providing consistent quantity-quality control across diverse scenes for robust deployment.
• It employs uniform Gaussian branching and opacity sparsity regularization for efficient, scene-agnostic Gaussian refinement and pruning.
• Control is managed by a single hyperparameter, showing consistent quality gains and reduced Gaussian counts across diverse scenes.

Consistent Quantity–Quality Control across Scenes for Deployment-Aware Gaussian Splatting

The paper introduces ControlGS, an innovative 3D Gaussian Splatting (3DGS) method that effectively addresses the complex trade-off between Gaussian quantity and rendering quality in the context of novel view synthesis (NVS). This paper presents a novel optimization strategy that enhances user control over this trade-off while maintaining substantial rendering quality across scenes of varied complexity without the need for scene-specific tuning.

3DGS has been recognized for its capability to balance rendering quality and real-time performance through its explicit anisotropic Gaussian projection approach. However, the explicit representation in 3DGS results in massive storage and computational burdens due to the management of millions of Gaussians. The fundamental challenge here is structural compression—achieving high rendering quality while reducing Gaussian numbers, which is crucial for deployment in environments with varied resource availability.

ControlGS tackles these issues by introducing key innovations. First, it implements a uniform Gaussian branching strategy that eschews traditional local heuristics for split/clone decisions, ensuring a consistent Gaussian density and effective usage globally across scenes. This is accomplished through a coarse-to-fine detail refinement progression, initialized with a sparse point cloud from Structure-from-Motion (SfM) processes and progressing to high-frequency detail capture through structured Gaussian splits.

Second, ControlGS incorporates a Gaussian atrophy mechanism enforced via opacity sparsity regularization. This mechanism provides an end-to-end, strength-controlled pruning methodology that corrects for Gaussian over-splitting and ensures efficient removal of redundant Gaussians without requiring complex hyperparameter tuning tied to individual scenes, thus reducing user intervention and increasing model deployment ease across different use-cases.

Importantly, this method consolidates control under a single user-specified hyperparameter, $\lambda_{\alpha}$, which governs the trade-off between Gaussian quantity and rendering quality. This control is robust across diverse scenes, spanning compact objects to expansive outdoor views and maintains predictability and consistency in performance, as evidenced across multiple datasets such as Mip-NeRF360 and Tanks and Temples.

The results demonstrate that ControlGS consistently improves rendering quality as measured by PSNR, SSIM, and LPIPS while simultaneously reducing Gaussian quantities used, outperforming methods like EAGLES and GoDe across a range of visual environments. This positions ControlGS as both a compact and high-quality alternative for Gaussian-based scene representation, showing potential to redefine deployment strategies in NVS applications.

Further work directions could involve integrating ControlGS with attribute compression techniques to further enhance model compactness, extending its applicability to dynamic environments and broader NVS tasks, and exploring real-time adaptation to different rendering quality expectations or hardware capabilities.

In conclusion, ControlGS exemplifies a marked advancement in 3DGS techniques, offering a practical, flexible, and efficient solution for overcoming the persistent challenges of structural compression in 3D representation. Its adaptability and control capabilities position it as a pivotal step towards more robust and deployment-ready 3DGS models.