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

• Traditional alpha blending makes Gaussian depth sorting the primary rendering bottleneck; Mobile-GS uses depth-aware weighting to blend without sorting, emphasizing near Gaussians over far ones.Source: §1 · p. 2
• 116 FPS on a Snapdragon 8 Gen 3 GPU, Mobile-GS delivers real-time mobile rendering with quality comparable to original 3DGS and surpasses previous lightweight approaches in speed and visual quality.Source: §1 · pp. 2–3
• Because unordered blending can cause transparency artifacts, Mobile-GS adds a neural enhancement conditioned on Gaussian attributes and viewing direction to improve view-dependent effects.Source: §1 · p. 2
• First-order spherical harmonics distillation reduces appearance parameters from the original third-order representation, guided by a pretrained teacher to improve rendering speed and reduce storage.Source: §1 · pp. 2–3
• Neural vector quantization groups Gaussian parameters with K-means and compresses distilled SH features with lightweight decoders to reduce storage costs and improve rendering speed.Source: §1 · p. 3
• Contribution-based pruning removes Gaussians with low opacity and scale, which the method treats as indicators of insignificant contribution, further reducing storage costs.Source: §1 · p. 3

3D Gaussian Splatting

• 3DGS represents scenes with anisotropic Gaussians and a tile-based differentiable rasterizer; prior work improves multiview constraints, Gaussian efficiency, and compression, but depth sorting burdens mobile real-time rendering.MVGS (Du et al., 2026) adds multi-view learning; Scaffold-GS (Lu et al., 2024) uses a hierarchical scaffold, and Mini-Splatting (Fang & Wang, 2024) combines pruning and densification.Source: §2 · p. 3

Order Independent Transparency

• Traditional transparency uses depth peeling or sorted A-buffers, while OIT approaches approximate compositing; recent sort-free 3DGS works, including SortFreeGS (Hou et al., 2025), target fast rendering without sorting.Source: §2 · p. 3
• Summation-based sort-free rendering can compute terms in any order because addition is commutative, but these methods can exceed edge-device storage and inference-delay limits.The passage describes summations over learnable parameters and depth, then notes that the methods are not directly suitable for edge devices.Source: §2 · p. 3

Gaussian Compression and Pruning

• Prior compression and pruning reduce redundancy but may degrade quality and retain storage overhead; Mobile-GS targets compact attributes while preserving essential Gaussian features for mobile deployment.Prior approaches include vector quantization and entropy encoding, LocoGS (Shin et al., 2025), and pruning methods LODGE (Kulhanek et al., 2025), MaskGaussian (Liu et al., 2025), and GaussianSpa (Zhang et al., 2025).Source: §2 · p. 4

3.1 DEPTH-AWARE ORDER-INDEPENDENT RENDERING ▶“

• Mobile-GS removes depth sorting through learnable view-conditioned weights, enabling parallel Gaussian accumulation while retaining depth- and scale-aware contributions.Inverse depth reduces distant Gaussians’ contributions, while larger-scale Gaussians receive greater weight. A lightweight MLP uses camera direction, Gaussian scale, rotation, and spherical harmonics to predict view-dependent opacity and attenuation, mitigating transparency artifacts in occluded or overlapping regions.Source: §3.1 · pp. 4–5

3.2 DISTILLATION AND QUANTIZATION ▶“

• Mobile-GS distills directional radiance into first-order spherical harmonics and compresses Gaussian attributes with sub-vector neural quantization and Huffman coding.It also uses scale-invariant teacher-depth supervision and reconstructs diffuse and view-dependent spherical-harmonic components with separate 16-bit MLPs, reducing per-Gaussian coefficient storage.Source: §3.2 · p. 6

3.3 CONTRIBUTION-BASED PRUNING

• Contribution-based pruning removes Gaussians that repeatedly show low opacity and small scale, using accumulated votes to avoid reacting to noisy early-training statistics.Candidates are identified using an opacity-and-scale quantile threshold and are permanently pruned when their accumulated votes exceed the threshold across pruning intervals.Source: §3.3 · pp. 6–7

3.4 IMPLEMENTATION

• Mobile-GS is trained for 60k iterations with custom CUDA kernels and deployed using Vulkan 2.0 for evaluation on mobile GPUs.Training uses Mini-Splatting (Fang & Wang, 2024) as the distillation teacher and launches neural vector quantization at iteration 35k.Source: §3.4 · p. 7

4.1 QUANTITATIVE AND QUALITATIVE RESULTS

• Mobile-GS produces sharp, consistent novel views comparable to 3DGS, with better view-dependent effects and strong results in scenes with complex geometry and lighting.The reported comparisons cover Mip-NeRF 360, Tank&Temples, and Deep Blending.Source: §4.1 · pp. 7–8, Table 1 · p. 8
• On Snapdragon 8 Gen 3, Mobile-GS outperforms the compared methods in rendering quality, speed, and storage cost, despite requiring more training time.The comparison quantizes 3DGS and Mini-Splatting for deployment; the authors attribute the gains to order-independent rendering, quantization, and pruning.Source: Table 2 · p. 8, §4.1 · pp. 8–9

4.2 ABLATION STUDY

• Ablations show that order-independent rendering improves speed, view-dependent enhancement preserves visual quality, and neural vector quantization and SH distillation reduce storage or rendering costs.Replacing order-independent rendering with alpha blending slightly improves PSNR but significantly reduces speed; omitting SH distillation reduces speed and increases storage, while omitting vector quantization increases storage.Source: Table 3 · p. 8, §4.2 · p. 9

4.1 QUANTITATIVE AND QUALITATIVE RESULTS

• Runtime analysis across four indoor and outdoor scenes finds that lightweight view-dependent MLPs add minimal overhead while preserving real-time efficiency and visual fidelity.The reported runtime includes all essential rendering-pipeline components.Source: Figure 6 · p. 9, §4.1 · p. 9

4.2 ABLATION STUDY

• Without view-dependent enhancement, transparency artifacts emerge around overlapping geometry and depth ambiguities; the neural strategy reduces them for more consistent, realistic views.It models relationships between Gaussian attributes and view-dependent lighting effects to address artifacts from order-independent rendering.Source: Figure 7 · p. 9, §4.2 · p. 9
• A codebook size of 2^10 balances storage and rendering quality in neural vector quantization; smaller codebooks sharply degrade PSNR, while larger ones require substantially more storage.The method uses K-means clustering to encode Gaussian parameters into a compact codebook.Source: Table 7 · p. 10, §4.2 · p. 10
• A threshold of 0.2 provides the best reported balance for contribution-based pruning, which combines opacity and scale to remove redundant Gaussians with minimal fidelity loss.Using either attribute alone degrades performance; the combined strategy can also further reduce Gaussian counts when integrated with existing pruning methods, without substantial performance degradation.Source: §4.2 · pp. 9–10, Table 4 · p. 10, Table 5 · p. 10

5 CONCLUSION

• Mobile-GS achieves quality comparable to original 3DGS, a smaller storage footprint, faster efficiency, and up to 127 FPS on a modern mobile GPU.Source: §5 · p. 10
• Its techniques include depth-aware order-independent rendering, neural view-dependent opacity enhancement, first-order spherical harmonics distillation, neural vector quantization, and contribution-based pruning.Together, these reduce computational overhead, memory, and storage while enabling high-quality novel view synthesis.Source: §5 · p. 10

A LLM USAGE STATEMENT

• The authors report using an LLM only to polish language, not to generate ideas, conduct analysis, or contribute scientific content.They state that the research design, implementation, and results are their original contributions.Source: A LLM USAGE STATEMENT · p. 15

B PRELIMINARY: 3D GAUSSIAN SPLATTING

• 3DGS represents scenes with anisotropic 3D Gaussians parameterized by position, covariance, opacity, and spherical-harmonic color coefficients.The covariance describes spatial extent and orientation and can be decomposed into scale and rotation.Source: B PRELIMINARY: 3D GAUSSIAN SPLATTING · p. 15
• Rendering projects each Gaussian onto the image plane, sorts Gaussians from near to far, and computes final pixel color through alpha blending.Source: B PRELIMINARY: 3D GAUSSIAN SPLATTING · p. 15
• Gaussian Splatting is described as differentiable, compact, and efficient for representing and rendering complex scenes, including in real-time applications and gradient-based neural-rendering optimization.Source: B PRELIMINARY: 3D GAUSSIAN SPLATTING · p. 15

Neural View-dependent Enhancement

• A three-layer MLP with 256, 128, and 64 neurons predicts opacity and an auxiliary feature, using Sigmoid and ReLU activations, and trains jointly with Gaussians for 30k iterations.The opacity output is constrained to [0, 1]; the auxiliary feature uses ReLU.Source: C ADDITIONAL IMPLEMENTATION DETAILS · p. 15

Neural Vector Quantization

• Gaussian attributes are split into five groups, K-means clustered, and Huffman-coded; per-group MLPs with one 64-neuron hidden layer refine quantized representations.The compact MLP design is intended to balance training efficiency and reconstruction accuracy.Source: C ADDITIONAL IMPLEMENTATION DETAILS · p. 15

Distribution-based Pruning

• Pruning uses Iprune = 1000, v = 0.6, and τ = 0.2, and runs only during the initial 25k iterations to remove marginally contributing Gaussians while preserving later-stage detail.The threshold selectively identifies redundant primitives; limiting pruning to early training helps avoid excessive removal when finer details become critical.Source: C ADDITIONAL IMPLEMENTATION DETAILS · p. 15

D.1 DIFFERENCE WITH SORTING-FREE METHODS

• Mobile-GS consistently outperforms prior sorting-free methods in PSNR, storage cost, and rendering FPS, with gains attributed to quantization, pruning, and view-dependent enhancement.SortFreeGS* denotes the quantized version of SortFreeGS; the original method has no quantization stage.Source: §D.1 · p. 16
• Mobile-GS uses a transmittance proxy enriched with view-dependent modulation, whereas GES follows a formulation similar to SortFreeGS.This formulation is described as better capturing the underlying 3D scene structure.Source: §D.1 · p. 16
• Mobile-GS weights each Gaussian using both depth and scale, unlike SortFreeGS, which modulates contribution using depth but omits scale.The method treats farther Gaussians as less contributive and larger Gaussians as providing more meaningful rendering evidence.Source: §D.1 · p. 16
• GES renders a depth image with conventional volume rendering, then filters distant Gaussians by comparing their depths with that image before sorting-free rendering.The two-stage pipeline requires precise depth rendering and increases computational load, making it poorly suited to mobile deployment.Source: §D.1 · p. 16
• Overlapping geometry creates order ambiguity for sorting-free methods, while SortFreeGS addresses it with additional spherical harmonics for view-dependent opacity at significant overhead.The passage characterizes that overhead as unfavorable for practical mobile use.Source: §D.1 · p. 16
• Mobile-GS models view-dependent effects with a learnable parameter ϕ predicted by a lightweight MLP conditioned on Gaussian attributes, avoiding prohibitive compute or memory burden.The discussion presents this as a way to achieve high-quality rendering while minimizing resource consumption and Gaussian parameter storage and maintaining real-time mobile rendering.Source: §D.1 · p. 16

D.2 DISCUSSION

• Depth-aware order-independent rendering replaces traditional alpha blending with a sorting-free scheme that substantially improves runtime efficiency.Neural view-dependent enhancements and spherical harmonics distillation help maintain visual fidelity even in complex scenes.Source: §D.2 · p. 16
• Neural vector quantization improves storage efficiency, while contribution-based pruning reduces the Gaussian representation for memory-constrained mobile deployment.Together, these strategies support deploying large-scale scene representations on mobile devices with limited memory.Source: §D.2 · pp. 16–17
• Mobile-GS consistently outperforms existing lightweight Gaussian Splatting methods across multiple benchmarks.The experiments highlight depth-aware order-independent rendering, neural view-dependent enhancement, spherical harmonics distillation, neural vector quantization, and contribution-based pruning.Source: §D.2 · pp. 16–17

D.3 LIMITATIONS

• Training remains computationally intensive, and desktop-GPU pretraining limits on-device retraining and real-time data acquisition.The costly components include spherical harmonics distillation, neural vector quantization, and neural view-dependent enhancement.Source: §D.3 · p. 17
• Mobile-GS is optimized per scene and requires retraining to generalize, limiting use in dynamic capture or unseen environments such as real-time AR reconstruction.Source: §D.3 · p. 17
• Neural vector quantization trades compression ratio against reconstruction quality; excessive quantization can shift colors or blur highly textured regions.The degradation particularly affects fine-grained appearance details.Source: §D.3 · p. 17

E ADDITIONAL QUALITATIVE AND QUANTITATIVE RESULTS

• Mobile-GS performs comparably to 3DGS, SortFreeGS, and Speedy-Splat across detailed per-scene results, surpassing them in several scenes.The comparison is reported in Tables 9–11; the supplied passage provides no specific metric values.Source: E ADDITIONAL QUALITATIVE AND… · p. 17
• Mobile-GS demonstrates superior novel-view rendering quality in the additional visual comparisons.The paper attributes the improvement to neural view-dependent enhancement, which better fits and adapts Gaussians to view-dependent effects.Source: E ADDITIONAL QUALITATIVE AND… · p. 17, Figure 8 · p. 18
• Neural view-dependent enhancement increases the representation capacity of 3D Gaussian properties to model view-dependent effects.The passage identifies this strategy as the primary reason for improved rendering results.Source: E ADDITIONAL QUALITATIVE AND… · p. 17
• Mobile-GS maintains high-quality novel-view synthesis while supporting real-time rendering on mobile devices.Source: E ADDITIONAL QUALITATIVE AND… · p. 17

F USRER STUDY

• The user study compared Mobile-GS with 3DGS (Kerbl et al., 2023) and LocoGS-S (Shin et al., 2025) on three public datasets, with quantized 3DGS for fairness.Thirty volunteers rated novel-view-synthesis videos rendered by the methods.Source: F USRER STUDY · p. 18, Figure 9 · p. 18
• Most of the 30 participants preferred Mobile-GS renderings, indicating higher visual quality than the alternatives in the study.The passage links this preference to mobile-tailored neural view-dependent enhancement and neural vector quantization; it notes that 3DGS often shows floaters and rendering artifacts.Source: F USRER STUDY · p. 18, Figure 9 · p. 18

G ADDITIONAL MOBILE TESTING

• The mobile evaluation reports cold-start FPS at launch and steady-state FPS after thermal equilibrium to assess performance changes over time.The paper attributes mobile FPS drops to thermal throttling, power limits, GPU clock downscaling, and NPU/CPU frequency limits.Source: G ADDITIONAL MOBILE TESTING · pp. 18–19
• Mobile-GS achieves 74 steady-state FPS on mobile, demonstrating effectiveness for real-time rendering despite performance changes over time associated with thermal throttling, power limits, and GPU, NPU, and CPU frequency limits.Source: Table 12 · p. 19, G ADDITIONAL MOBILE TESTING · p. 19
• Mobile-GS achieves the lowest power consumption among Mobile-GS, 3DGS*, and SortFreeGS* on the mobile device.3DGS* and SortFreeGS* are quantized variants adapted for the Snapdragon 8 Gen 3 mobile GPU; power was measured for Vulkan operators and the MLP on Mip-NeRF 360.Source: Table 13 · p. 19, G ADDITIONAL MOBILE TESTING · p. 19