Xiaobiao Du

N[𝟙​(di<ds​(𝐱^)+ϵ)]​αi​(𝐱^)W_{G}(\hat{\mathbf{x}})=\sum_{i=1}^{N}[\mathbbm{1}(d_{i}<d_{s}(\hat{\mathbf{x}})+\epsilon)]\alpha_{i}(\hat{\mathbf{x}}) 27.02 29.4 MB 24 Ours 𝐂=(1−T)​∑i=1𝒩𝐜i​αi​wi∑i=1𝒩αi​wi+T​𝐜b​g\mathbf{C}=\left(1-T\right)\frac{\sum_{i=1}^{\mathcal{N}}\mathbf{c}_{i}\alpha_{i}w_{i}}{\sum_{i=1}^{\mathcal{N}}\alpha_{i}w_{i}}+T\mathbf{c}_{bg} wi=ϕi2+ϕidi2+e​x​p​(sm​a​xdi)w_{i}=\phi^{2}_{i}+\frac{\phi_{i}}{d_{i}^{2}}+exp(\frac{s_{max}}{d_{i}}) 27.12 4.6 MB 127 D.2 Discussion Mobile-GS is a Gaussian-based method that can achieve real-time rendering on mobile and resource-constrained platforms without significantly sacrificing rendering quality. The proposed depth-aware order-independent rendering replaces traditional alpha blending with a sorting-free scheme, substantially improving runtime efficiency. Combined with neural view-dependent enhancements and spherical harmonics distillation, our approach maintains visual fidelity even under complex scenes. To address memory limitations, a neural vector quantization strategy is employed, improving storage efficiency and enabling large-scale scene representations to be deployed on mobile devices with limited memory. Experimental results demonstrate that Mobile-GS achieves a compelling balance among rendering speed, storage footprint, and visual quality. It consistently outperforms existing lightweight Gaussian Splatting methods across multiple benchmarks, highlighting the effectiveness of our proposed components, including depth-aware order-independent rendering, neural view-dependent enhancement, spherical harmonics distillation, neural vector quantization, and contribution-based pruning. D.3 Limitations Despite its advantages, Mobile-GS contains several limitations: (1) Training Cost and Complexity: Although inference is fast, training Mobile-GS remains computationally intensive due to the proposed components (e.g., spherical harmonics distillation, neural vector quantization, neural view-dependent enhancement). Additionally, the model requires pretraining on desktop GPUs before mobile deployment, limiting its accessibility for real-time data acquisition and retraining on the device. (2) Scene Generalization: While Mobile-GS performs well on standard benchmarks, it is optimized per-scene and does not generalize across scenes without retraining. This limits its immediate usage in applications requiring dynamic scene capture or rendering in unseen environments, such as real-time AR reconstruction. (3) Quantization Degradation: Although the proposed neural vector quantization is highly effective in compressing Gaussian attributes, there still remains a trade-off between compression ratio and reconstruction quality, especially for fine-grained appearance details. Excessive quantization may introduce minor color shifts or blurring artifacts in highly textured regions. Table 9: Per-scene PSNR results of state-of-the-art novel view synthesis methods on Mip-NeRF 360 dataset (Barron et al., 2022). The best results are highlighted. Method bicycle garden stump flowers treehill counter kitchen room bonsai, D scene structure. The Gaussian weight computation also differs substantially across these approaches. SortFreeGS leverages the Gaussian depth to modulate its contribution but does not account for the Gaussian scale, which we find to be critical. GES, on the other hand, relies on a two-stage rendering. It first renders a depth image using conventional volume rendering and then filters out distant Gaussians by comparing their depths against the rendered depth map for later sorting-free rendering. This two-stage rendering pipeline relies on precise depth rendering and increases computational load, so it is not well-suited for mobile deployment. In contrast, Mobile-GS exploits both depth and scale attributes of each Gaussian to compute an importance weight, reflecting the intuition that farther Gaussians should have lower contribution, while larger Gaussians typically provide more meaningful rendering evidence. Theoretically, A key challenge for sorting-free methods is the potential order ambiguity in regions where geometry overlaps. SortFreeGS attempts to address this by introducing additional spherical harmonics parameters to model view-dependent opacity. However, this design incurs significant overhead and is unfavorable for practical mobile usage. Our Mobile-GS resolves this limitation by enhancing the view-dependent effect through a learnable parameter ϕ\phi, predicted by a lightweight MLP conditioned on Gaussian attributes. This formulation achieves high-quality rendering without introducing a prohibitive computational or memory burden. Overall, Mobile-GS is carefully tailored to minimize resource consumption, reduce Gaussian parameter storage, and maintain real-time rendering performance on mobile hardware. Table 8: Comparison with different sorting-free methods. SortFreeGS* means its quantized version. We report metrics on the Mip-NeRF 360 dataset for the mobile equipped with the Snapdragon 8 Gen 3 GPU. FPS* means the rendering speed on the mobile. Method Rendering Weighting PSNR ↑\uparrow Storage ↓\downarrow FPS*↑\uparrow SortFreeGS* 𝐂=cb​g​wb​g+∑i=1𝒩ci​αi​w​(di)wb​g+∑i=1𝒩αi​w​(di)\mathbf{C}=\frac{c_{bg}w_{bg}+\sum_{i=1}^{\mathcal{N}}c_{i}\alpha_{i}w(d_{i})}{w_{bg}+\sum_{i=1}^{\mathcal{N}}\alpha_{i}w(d_{i})} w​(di)=exp⁡(−σ​diβ)w(d_{i})=\exp\!\left(-\sigma d_{i}^{\beta}\right) 26.74 64.3 MB 18 GES 𝐂=Cs​Ws+CGWs+WG\mathbf{C}=\frac{C_{s}W_{s}+C_{G}}{W_{s}+W_{G}} WG​(𝐱^)=∑i=