Mobile-GS: Real-time Gaussian Splatting for Mobile Devices

Abstract: 3D Gaussian Splatting (3DGS) has emerged as a powerful representation for high-quality rendering across a wide range of applications.However, its high computational demands and large storage costs pose significant challenges for deployment on mobile devices. In this work, we propose a mobile-tailored real-time Gaussian Splatting method, dubbed Mobile-GS, enabling efficient inference of Gaussian Splatting on edge devices. Specifically, we first identify alpha blending as the primary computational bottleneck, since it relies on the time-consuming Gaussian depth sorting process. To solve this issue, we propose a depth-aware order-independent rendering scheme that eliminates the need for sorting, thereby substantially accelerating rendering. Although this order-independent rendering improves rendering speed, it may introduce transparency artifacts in regions with overlapping geometry due to the scarcity of rendering order. To address this problem, we propose a neural view-dependent enhancement strategy, enabling more accurate modeling of view-dependent effects conditioned on viewing direction, 3D Gaussian geometry, and appearance attributes. In this way, Mobile-GS can achieve both high-quality and real-time rendering. Furthermore, to facilitate deployment on memory-constrained mobile platforms, we also introduce first-order spherical harmonics distillation, a neural vector quantization technique, and a contribution-based pruning strategy to reduce the number of Gaussian primitives and compress the 3D Gaussian representation with the assistance of neural networks. Extensive experiments demonstrate that our proposed Mobile-GS achieves real-time rendering and compact model size while preserving high visual quality, making it well-suited for mobile applications.

• The paper proposes an order-independent rendering method that eliminates alpha blending sorting, significantly accelerating real-time 3D rendering on mobile devices.
• It integrates a neural view-dependent opacity enhancement using a lightweight MLP to optimize and reduce transparency artifacts during compositing.
• The framework employs aggressive model compression via SH distillation, neural vector quantization, and contribution-based pruning to lower storage and computational overhead.

Real-time 3D Gaussian Splatting on Mobile Devices with Mobile-GS

Introduction

The paper "Mobile-GS: Real-time Gaussian Splatting for Mobile Devices" (2603.11531) presents a comprehensive system for real-time 3D scene rendering on mobile devices, using an optimized and compressed form of 3D Gaussian Splatting (3DGS). While 3DGS achieves photorealistic novel view synthesis via anisotropic Gaussian primitives, existing approaches incur excessive computational and memory overhead, particularly due to the alpha blending depth-sorting bottleneck and large parameter storage. Mobile-GS addresses these challenges by proposing an integrated framework with order-independent rendering, neural view-dependent enhancement, aggressive quantization, and contribution-based pruning, facilitating deployment on edge platforms such as smartphones and AR headsets.

Depth-aware Order-independent Rendering

Traditional 3DGS relies on alpha blending, necessitating strict depth ordering of Gaussian primitives during compositing—an operation fundamentally ill-suited for parallelism on mobile hardware and accounting for the majority of the runtime (up to half the render time). Mobile-GS circumvents this by introducing a depth-aware, order-independent rendering formulation where per-Gaussian contributions are modulated via learnable, view-conditioned weights based on position, scale, and relative camera direction.

This strategy allows rendering with a single pass over the set of Gaussians and eliminates sorting altogether, yielding a several-fold inference speed-up. Importantly, this order-independent mechanism preserves most of the depth discrimination required for correct foreground/background separation by emphasizing near Gaussians and downweighting distant ones.

Neural View-dependent Opacity Enhancement

Order-independent blending introduces potential transparency artifacts, particularly for overlapping and occluded geometry, which canonical 3DGS addresses with explicit sorting. To mitigate this, Mobile-GS augments each Gaussian with a learned, view-dependent opacity scalar. This is regressed by a lightweight MLP conditioned on Gaussian geometry (position, scale, rotation), spherical harmonics encoding appearance, and camera-primitive direction.

The resulting pipeline adaptively governs the per-view visibility of Gaussians, attenuating transparency errors and improving rendition of view-dependent effects (e.g., specularities). The ablation demonstrates severe quality degradation without this component, affirming its necessity.

Aggressive Model Compression: SH Distillation and Neural Vector Quantization

To meet the stringent memory and bandwidth constraints of mobile hardware, Mobile-GS aggressively compresses the Gaussian parameterization:

• First-order SH Distillation: Instead of storing high-dimensional third-order SH appearance coefficients, a distillation pipeline is enforced to regress first-order coefficients under supervision from a stronger teacher model (e.g., Mini-Splatting). This reduces the per-Gaussian storage burden and inference complexity with minimal PSNR drop.
• Neural Vector Quantization: Gaussian parameters are partitioned using K-means into sub-vectors, each encoded with a dedicated codebook, followed by Huffman entropy coding for further redundancy reduction. At inference, compact MLPs reconstruct the SH features, further minimizing overhead.

Empirical results show a drastic drop in Gaussian storage—down to ∼4.6 MB per scene—exceeding the compression achieved by methods such as LocoGS-S, Speedy-Splat, and others, with only negligible PSNR loss.

Contribution-based Pruning for Structural Compactness

Redundant or low-contribution Gaussians are pruned dynamically during training according to a joint statistic over per-Gaussian opacity and maximal spatial extent. An accumulated voting scheme ensures robustness against noisy gradient fluctuations, with ablation indicating significant performance drops if pruning relies solely on either opacity or scale.

The pruning threshold is tuned to achieve a balanced trade-off between rendering quality and primitive count, with Mobile-GS at default settings maintaining state-of-the-art visual fidelity while operating with about 0.47 million Gaussians per scene—a substantial reduction relative to full-capacity 3DGS and competitive lightweight variants.

Experimental Evaluation

Benchmarking on standard datasets (Mip-NeRF 360, Tank&Temples, Deep Blending) demonstrates that Mobile-GS exhibits rendering quality on par with or superior to conventional 3DGS and state-of-the-art compact methods, while consuming dramatically less storage and delivering real-time framerates on mobile hardware:

• Rendering Speed: 1098 FPS (RTX 3090 Ti), up to 127 FPS on a Snapdragon 8 Gen 3 GPU, and 74 FPS steady-state after thermal equilibrium.
• Storage: ∼4.6 MB per scene, lowest among all high-quality contenders.
• Quality: PSNR 27.12–29.93, with competitive SSIM and LPIPS.

Power profiling reveals Mobile-GS as the lowest-power design—even under the constraints of mobile GPUs running Vulkan 2.0—underscoring the practical impact of the design decisions.

Analysis and Limitations

Mobile-GS positions itself as a unified, practical solution for Gaussian-based mobile rendering. Its integrated optimizations—order-independent blending, neural view-dependent modulation, SH compression, and deep pruning—collectively close the efficiency gap that has prevented prior 3DGS systems from practical edge deployment.

However, limitations persist. Training remains computationally expensive, requiring desktop resources and scene-specific fine-tuning that preclude direct, on-device training or transfer to arbitrary new scenes. Additionally, the compression pipeline, while highly effective, introduces a hard trade-off between storage and fine detail; excessive quantization may lead to blurring or color shifts in highly textured regions, especially at low codebook sizes. Generalization across diverse scenes without retraining is also unaddressed, as the model is not designed for dynamic scene adaptation.

Implications and Future Directions

The Mobile-GS framework opens the path for photorealistic scene capture and rendering on AR/VR headsets and mobile phones with real-time performance, enabling applications in mobile gaming, telepresence, and in-the-wild mixed reality. The adoption of learned view-dependent opacity and compressed SH representations points toward a general paradigm for mobile-friendly neural graphics.

Further progress is likely to center on scene-generalizing models, on-device continual adaptation methods, and task- or region-aware allocation of quantization capacity. Additionally, research into training-time and hardware-level co-design could further reduce mobile inference time and energy cost, critical for sustained usage on thermally constrained devices.

Conclusion

Mobile-GS constitutes a significant advancement in the deployment of 3D Gaussian Splatting for mobile and embedded platforms. By eliminating the alpha-sorting bottleneck and integrating neural modulation, model distillation, vector quantization, and adaptive pruning, Mobile-GS achieves real-time rendering under tight memory and energy budgets, without significant loss in rendering fidelity. The design and results of Mobile-GS delineate key techniques and trade-offs required for practical neural scene representation, establishing a foundation for future work in high-fidelity, real-time, and resource-efficient 3D scene rendering.