Monte Carlo Energy Aggregation for Mobile 3D Gaussian Splatting

Conference: ECCV 2026
arXiv: 2606.30017
Code: https://xiaobiaodu.github.io/flux-gs-project/
Area: 3D Vision / Model Compression
Keywords: 3D Gaussian Splatting, Mobile Rendering, Spherical Harmonics Compression, Multi-view Consistency, Neural Rendering

TL;DR

Flux-GS compresses 3rd-order Spherical Harmonics (SH) into a 1st-order representation via Monte Carlo specular energy aggregation. Complemented by an attribute-conditioned enhancement module with zero inference overhead and a multi-view alpha-weighted densification/pruning strategy, it achieves real-time, high-fidelity novel view synthesis at 130+ FPS on a mobile Snapdragon 8 Gen 3. It also compresses storage to 2-5 MB and reduces training time from Mobile-GS's 80+ minutes to 11 minutes.

Background & Motivation

Background: 3D Gaussian Splatting (3DGS) has become the mainstream paradigm for 3D reconstruction and rendering, achieving high-fidelity real-time novel view synthesis on desktop GPUs due to its explicit anisotropic Gaussian representation. However, each Gaussian primitive requires 48 3rd-order Spherical Harmonics (SH) coefficients to model view-dependent radiance. When millions of Gaussian primitives accumulate, the storage and VRAM bandwidth requirements become prohibitive for mobile devices.

Limitations of Prior Work: Existing lightweighting efforts proceed along two lines: (1) SH compression and distillation: Mobile-GS transfers 3rd-order SH to 1st-order via teacher-student distillation, but the distillation process suffers from extremely high training costs (80+ minutes), and its view-dependent MLP enhancement module still requires online computation during inference, slowing down rendering. (2) Gaussian pruning: most methods prune based on single-view gradients or importance scores, lacking multi-view structural awareness, which easily leads to over-densification and redundant Gaussian primitives, dragging down training efficiency and expanding storage.

Key Challenge: High-fidelity rendering demands high-order SH to capture complex specular reflections and view-dependent effects, yet the storage (48 floating-point coefficients per Gaussian) and bandwidth overhead of high-order SH fundamentally conflict with the stringent resource constraints of mobile devices. Concurrently, single-view-driven densification strategies aggressively expand the primitive count in pursuit of reconstruction accuracy, forming another contradiction with the compact representation required by mobile devices.

Goal: To compress the specular energy of 3rd-order SH into a compact 1st-order subspace without relying on expensive distillation or pre-training, while steering the addition and removal of Gaussian primitives using multi-view consistency, allowing the model to simultaneously achieve high frame rates, low storage, and rapid training on mobile devices.

Key Insight: Instead of performing pixel-wise radiance residual aggregation (which would lose angular gradient information), Monte Carlo sampling is utilized to extract the first-order directional moments (energy magnitude + dominant direction) of the specular residuals, encoding sparse specular peak information into a compact latent space. Rather than relying on single-view gradient-driven densification, multi-view alpha-weighted error accumulation is used to identify the Gaussian primitives that truly need to be added or removed.

Core Idea: Compress high-order SH view-dependent signals into compact "energy + direction" descriptors, which are then decoded into 1st-order SH offsets once before inference using a geometrically-conditioned MLP—guaranteeing zero extra overhead during inference; meanwhile, multi-view reconstruction errors backwardly guide which Gaussians should split or be pruned.

Method

Overall Architecture

The training of Flux-GS is divided into two phases. The first 3k iterations use standard 3rd-order SH training to establish a high-frequency representation; subsequently, a Monte Carlo Specular Energy Aggregator compresses the 3rd-order SH into 1st-order SH in a one-time operation, initializing the attribute-conditioned SH enhancement module to continuously refine the 1st-order representation during the remaining 27k iterations. During final inference, the SH offsets predicted by the enhancement module are statically baked into the Gaussian parameters, requiring no extra MLP inference in the rendering pipeline.

At the core of the framework is a pipeline of three contributing modules: the Monte Carlo Specular Energy Aggregator is responsible for "compression" (fidelity-preserving compression from 3rd-order to 1st-order), the attribute-conditioned SH enhancement module handles "compensation" (restoring high-frequency details lost in compression), and the multi-view alpha-weighted densification and pruning controls "quantity/quality" (managing the number and quality of Gaussian primitives).

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: Multi-view images + SfM point cloud"] --> B["Phase I (0-3k iterations)<br/>Standard 3DGS training<br/>Full 3rd-order SH optimization"]
    B --> C["Monte Carlo Specular Energy Aggregation<br/>3rd-order SH → 1st-order SH<br/>Compressed to Emag + Edir latent variables"]
    C --> D["Attribute-Conditioned SH Enhancement<br/>MLP predicts 1st-order SH offset Δc<br/>Statically baked before inference, zero inference overhead"]
    D --> E["Phase II (3k-30k iterations)<br/>Joint optimization of 1st-order SH + Δc"]
    E --> F["Multi-view Alpha Densification & Pruning<br/>Hierarchical camera sampling = Multi-view error maps<br/>= alpha-weighted importance/pruning scores"]
    F --> G["Output: Compact 1st-order SH Gaussian field<br/>WebGL mobile real-time rendering"]

Key Designs

1. Monte Carlo Specular Energy Aggregator: Compressing 3rd-order SH specular information into a compact latent space of "energy magnitude + direction"

High-order SH utilizes numerous coefficients to cancel out non-specular ringing artifacts in order to express a sparse specular peak—this is highly inefficient. Simply discarding high-order coefficients loses crucial view-dependent details. However, performing a straightforward spherical integration on the residuals results in zero integrals for high-order terms due to SH orthogonality, rendering the effort futile. This paper's insight is that the perceptual impact of specular reflection is fundamentally determined by "how bright it is" and "in which direction it lies." These two quantities can be extracted by performing Monte Carlo directional moment estimation on the high-order residuals.

Specifically, \(K=2048\) directions \(\mathbf{d}_k\) are rendered uniformly on the unit sphere. The residual \(c_{res}(\mathbf{d}_k) = c_i(\mathbf{d}_k) - c_i^2(\mathbf{d}_k)\) is computed between the 3rd-order SH radiance \(c_i(\mathbf{d}_k)\) and the 2nd-order SH radiance \(c_i^2(\mathbf{d}_k)\). Then, the energy magnitude \(E_{mag} = \frac{1}{K}\sum \max(0, c_{res}(\mathbf{d}_k))\) and direction \(E_{dir} = \frac{1}{K}\sum \mathbf{d}_k \otimes \max(0, c_{res}(\mathbf{d}_k))\) (outer product) are aggregated. These two quantities encode the overall brightness and dominant incident direction of the specular residual on each Gaussian primitive, with dimensions far lower than the original 48 SH coefficients. Subsequently, a lightweight MLP \(\Psi\) maps \((E_{mag}, E_{dir}, f(\hat{\mu}))\) to the 1st-order SH coefficients \(c'\) (conditioned on the normalized Gaussian position \(\hat{\mu}\)). The entire process is executed only once at the 3k-th iteration, after which the model only maintains 1st-order parameters.

Compared with the distillation scheme of Mobile-GS, MC-SEA does not require training a high-order teacher model or performing scene-by-scene distillation, slashing training time from 80+ minutes to 11 minutes.

2. Attribute-Conditioned SH Enhancement Module: Predicting SH offsets using Gaussian geometric attributes to achieve zero inference overhead via pre-inference static baking

Compressing from 3rd-order to 1st-order inevitably loses color fidelity and scene structure information. The authors found that these discarded residual energies are highly correlated with the intrinsic attributes of the Gaussian primitives (spatial position, anisotropic scale, opacity, rotation quaternion). Based on this, a lightweight MLP \(\Phi\) is designed, taking the concatenated vector of normalized 1st-order SH \(\hat{c}'\), opacity \(o\), normalized scale \(\hat{s}\), position \(\mu\), and rotation \(r\) as input, and outputting the residual offset \(\Delta c\) of the 1st-order SH: \(c^{out} = c' + \Delta c\).

Two key details ensure practicality and stability: (1) The last layer of \(\Phi\) is zero-initialized, allowing the model to naturally fall back to standard 3DGS behavior in the early stage of training, and then smoothly learn the residuals. (2) \(\Phi\) does not depend on the camera viewing direction; thus, \(\Delta c\) only needs to be decoded once before inference and statically baked into the Gaussian parameters, requiring no MLP forward passes during inference—genuinely achieving zero inference overhead. The corresponding module in Mobile-GS relies on online viewing direction inference, which slows down rendering.

3. Multi-View Alpha-Weighted Densification and Pruning: Guiding the addition and removal of Gaussians backwardly via multi-view reconstruction errors instead of single-view gradient-driven strategies

Standard 3DGS densification only considers the single-view position gradient \(\|\nabla_{\mu_i}\mathcal{L}\|_2 > \tau\), which easily overfits a specific viewing angle and generates many redundant Gaussians. Flux-GS takes a three-step approach:

Hierarchical Camera Sampling: First estimate the scene center \(\mathbf{x}^*\) (by solving the least-squares intersection of all camera rays). Convert camera positions to spherical coordinates and bin them by azimuth and elevation. Uniformly sample one representative view from each non-empty bin, and finally randomly truncate to a budget \(K\) to obtain an angularly uniform camera subset \(\mathcal{C}\).

Photometric Loss Guidance: For each view in \(\mathcal{C}\), render and compute the pixel-wise hybrid loss \(\mathcal{L} = (1-\lambda)\mathcal{L}_1 + \lambda\mathcal{L}_{DSSIM}\). Binarize using thresholds \(\tau^+\) and \(\tau^-\) to obtain pixel masks \(M^+\) ("poor reconstruction") and \(M^-\) ("good reconstruction").

Alpha-Weighted Error Accumulation: A customized CUDA kernel is implemented to sum the alpha contributions of each Gaussian \(i\) in the \(M^+\) region to obtain an importance score \(S_i^+\), and sum the alpha-weighted loss in the \(M^-\) region to obtain a pruning score \(S_i^-\). The final clone/split mask intersects the baseline gradient condition with an importance quantile filter: \(\mathcal{M}_{clone} = \mathcal{M}_{clone}^{base} \land (S^+ > Q_{\tau^+}(S^+))\), ensuring that only Gaussians consistently contributing to high-error regions across multiple views are densified. The pruning mask is defined as \(\mathcal{M}_{prune} = (o_i < o_{min}) \land (S^- > Q_{\tau^-}(S^-))\), which retains standard low-opacity pruning while introducing multi-view redundancy judgment to realize "moderate pruning" rather than aggressive deletion.

A Complete Example: From 3rd-Order SH Training to 1st-Order SH Mobile Inference

Taking a Mip-NeRF 360 indoor scene (e.g., kitchen) as an example to walk through the entire pipeline: The initial SfM point cloud provides around 5k initial Gaussian positions. Flux-GS trains with the standard 3DGS pipeline for the first 3k iterations—each Gaussian maintains 48 3rd-order SH coefficients. Accompanied by multi-view alpha-weighted densification, the Gaussian count gradually expands from 5k to approximately 360k. At this point, PSNR is around 29.5 and storage is about 150 MB (dominated by 3rd-order SH), rendering it completely unrunnable on mobile devices.

At the 3k-th iteration, MC-SEA compression is executed: For each of the 360k Gaussians, 2048 directions are sampled on the sphere to compute the residual radiance \(c_{res}(\mathbf{d}_k)\) between 3rd-order and 2nd-order SH. This is aggregated into \(E_{mag} \in \mathbb{R}^3\) and \(E_{dir} \in \mathbb{R}^{3 \times 3}\), and then decoded in a one-time forward pass by MLP \(\Psi\) into 12 1st-order SH coefficients (4 coefficients/channel \(\times\) 3 channels). The 3rd-order SH coefficients are subsequently discarded, dropping storage from 150 MB to about 20 MB. The PSNR drops to approximately 26.6 (a loss of ~2.9 dB), primarily losing specular details.

Immediately afterwards, the attribute-conditioned SH enhancement module \(\Phi\) (a 4-layer MLP, with the final layer zero-initialized) is initialized. For each Gaussian, it takes position \(\mu\), scale \(\hat{s}\), rotation \(r\), opacity \(o\), and the freshly decoded 1st-order SH \(\hat{c}'\) as input to predict the residual \(\Delta c\), which is added to \(c'\). Due to the zero initialization, \(\Delta c = 0\) at startup, enabling a smooth transition. Over the next 27k iterations, \(\Phi\) gradually learns to compensate for the high-frequency details lost in compression using Gaussian geometric attributes, allowing the PSNR to recover from 26.6 to 27.02. Multi-view alpha densification and pruning run continuously during this period—each round computes pixel-wise loss masks from 6 hierarchically sampled views to accumulate alpha-weighted scores, performing clone/split only on Gaussians in "perpetually high multi-view error" regions (\(S^+ > Q_{0.6}\)), and pruning only "low-error and low-opacity" Gaussians (\(o < o_{min}\) and \(S^- > Q_{0.1}\)). The final Gaussian count stabilizes at around 366k, and storage is compressed to 3.5 MB after quantization.

Before inference, \(\Phi\) performs a single forward pass for all 366k Gaussians, and the predicted \(\Delta c\) is statically written back into each Gaussian's SH coefficients. During inference, the WebGL rendering pipeline only reads the baked 1st-order SH + geometric parameters for rasterization—bypassing any MLP computations entirely—and achieves a mobile frame rate of 139 FPS.

Loss & Training

The training target is consistent with standard 3DGS, adopting a hybrid photometric loss \(\mathcal{L} = (1-\lambda)\mathcal{L}_1 + \lambda\mathcal{L}_{DSSIM}\) while jointly optimizing Gaussian geometric parameters and compressed SH features. Out of 30k total iterations, the first 3k iterations use full 3rd-order SH to establish a high-frequency base. A one-time MC-SEA compression is executed at the 3k-th iteration, after which 3rd-order SH is disabled, and only 1st-order SH + the attribute-conditioned enhancement module are maintained. No pre-training or distillation is required. Hyperparameters for MC-SEA: spherical sample size \(K=2048\) (ablation shows PSNR saturates after 2048); the two MLPs are a single-hidden-layer MLP with 64 neurons and a 4-layer (128, 64, 32, 12) ReLU MLP, respectively. For multi-view densification and pruning: 6 sampled views, \(\tau^+ = 0.1\), \(\tau^- = 0.01\), \(Q_{\tau^+} = 0.6\), and \(Q_{\tau^-} = 0.1\). The quantization scheme follows that of Mobile-GS.

Key Experimental Results

Main Results

Flux-GS was evaluated on the Mip-NeRF 360 dataset using a Snapdragon 8 Gen 3 mobile GPU and an RTX 4090 desktop GPU. The main table (Table 1) is presented below, reporting indoor and outdoor scenes separately.

Method	PSNR↑ (Indoor)	#G×10⁶↓ (Indoor)	Storage MB↓ (Indoor)	FPS↑ (Indoor)	Train min↓ (Indoor)	PSNR↑ (Outdoor)	#G×10⁶↓ (Outdoor)	Storage MB↓ (Outdoor)	FPS↑ (Outdoor)	Train min↓ (Outdoor)
3DGS	30.41	1.45	478	-	27	24.61	3.14	1361	-	36
3DGS* (Quant.)	30.04	1.45	46	11	27	24.39	3.14	85	5	36
C3DGS	30.01	0.75	21	18	31	24.38	0.91	34	13	45
Mobile-GS	30.37	0.38	3.5	131	86	24.51	0.58	5.5	114	136
Mobile-GS* (No MLP)	29.58	0.38	3.4	142	64	23.74	0.58	5.4	128	114
Ours	30.22	0.22	2.1	147	11	24.45	0.48	4.6	132	11

Flux-GS achieves the highest FPS (147/132), the lowest Gaussian count (0.22M/0.48M), and the smallest storage (2.1/4.6 MB) across both indoor and outdoor scenes. The PSNR gap compared to Mobile-GS is within 0.15/0.06 dB, while its training time is only 1/8 to 1/12 of the latter.

Evaluation on the Tanks&Temples and Deep Blending datasets (Table 2):

Method	T&T PSNR↑	T&T SSIM↑	T&T LPIPS↓	T&T Storage↓	T&T FPS↑	DB PSNR↑	DB SSIM↑	DB LPIPS↓	DB Storage↓	DB FPS↑
3DGS	23.14	0.841	0.183	358.7	-	29.41	0.903	0.243	697.3	-
C3DGS	23.32	0.831	0.202	21.8	15	29.73	0.900	0.258	24.7	16
Mobile-GS*	22.41	0.825	0.227	2.5	129	29.14	0.895	0.281	4.5	146
Flux-GS	23.27	0.827	0.221	2.4	137	29.73	0.898	0.275	1.9	158

Flux-GS on T&T and DB obtains the highest FPS and smallest storage, and its PSNR/SSIM/LPIPS scores are close to C3DGS and significantly outperform Mobile-GS* without MLP.

Ablation Study

Ablation study on Mip-NeRF 360 (Table 3):

Config	PSNR↑	Storage MB↓	Peak VRAM↓	FPS↑	#Points×10⁶↓	Train min↓
Flux-GS (full)	27.02	3.5	388	139	0.36	11
w/o MC-SEA	26.64	3.5	388	139	0.36	11
w/o Δc (SH Enhancement Offset)	26.71	3.5	388	139	0.36	9
w/o Multi-view Densification	27.14	16	1591	24	1.37	18
w/o Multi-view Pruning	27.09	4.8	416	119	0.59	13

Removing multi-view densification (falling back to standard single-view gradient densification) causes the Gaussian count to skyrocket from 0.36M to 1.37M (a 3.8x increase), the storage to balloon from 3.5 MB to 16 MB (a 4.6x increase), and the FPS to plunge from 139 to 24—making it the most significant single contributing module. Removing MC-SEA drops the PSNR by 0.38, and removing \(\Delta c\) drops the PSNR by 0.31, proving their complementary nature.

Key Findings

• Multi-view densification is the soul of efficiency: Reverting to single-view gradient densification balloons the Gaussian count by nearly 4 times, indicating that the multi-view alpha-weighted constraint accurately suppresses overfitting to single views, which is the foundational reason Flux-GS maintains image quality with minimal primitives.
• MC-SEA and SH enhancement are complementary: Removing MC-SEA (-0.38 PSNR) or Removing \(\Delta c\) (-0.31 PSNR) individually yields lower scores than the full model, showing that energy aggregation preserves the dominant low-frequency information, while the enhancement module compensates for the geometry-related high-frequency residuals.
• The sample point count K=2048 is the sweet spot: As \(K\) increases from 64 to 2048, PSNR rises monotonically, saturating at \(K=4096\) while computational overhead grows linearly—concluding 2048 as the optimal trade-off between quality and efficiency.
• The multi-view camera count of 6 is optimal: Increasing from 2 to 6 views improves PSNR while significantly reducing the Gaussian count, with PSNR stabilizing beyond 6—suggesting that a 6-view multi-view constraint is sufficient to distinguish between important and redundant Gaussians.
• Quantitative analysis: A larger densification quantile \(Q_{\tau^+}\) results in fewer Gaussians being split/cloned, gradually reducing PSNR (27.02 at 0.6 vs. 26.91 at 0.9). A smaller pruning quantile \(Q_{\tau^-}\) leads to more aggressive pruning, dropping PSNR (26.81 at 0.01 vs. 27.02 at 0.1), marking 0.1 as the sweet spot.
• User study: A blind study with 30 volunteers across Mip-NeRF 360, T&T, and DB datasets shows that Flux-GS's subjective image quality is significantly preferred over Mobile-GS, Speedy-Splat, and the quantized version of 3DGS.

Highlights & Insights

• The perspective of replacing "residual aggregation" with "directional moments" is highly inspiring: The conventional approach to order reduction is to directly project the pixel residuals, which is blocked by the zero-integration issue (SH orthogonality). Instead, estimating the "first-order moments" of the residuals—energy magnitude and dominant direction—bypasses this mathematical hurdle while capturing the two most fundamental physical variables of specularity. This philosophy of "compressing the moments of the signal rather than the signal itself" can be extended to other dimension-reduction scenarios requiring structural protection (e.g., view-dependent color dimension reduction in NeRF, environment map probe compression).
• "Pre-inference static baking" is the golden rule for mobile deployment: The training-time computational overhead of the attribute-conditioned SH enhancement module is acceptable, but the authors strictly ensured it has no dependency on the viewing direction, running it only once before inference to bake it into the parameters. This FPS advantage over Mobile-GS's online MLP scheme fully instantiates the design philosophy of "costly training, lightweight inference." This has direct reference value for designing other mobile-friendly models.
• Alpha-weighting is more suitable for Gaussian pruning than gradient-weighting: Gradients reflect "how much a Gaussian has been optimized," whereas alpha reflects "how much a Gaussian contributes to the final pixel"—the latter being a more direct metric of Gaussian importance. With multi-view aggregation, it effectively distinguishes Gaussians that are "important across multiple views" from those that are "only important for single-view artifacts," offering a more natural pruning criterion than other importance scoring methods (e.g., transmittance, Hessian sensitivity).
• Lightweight design of hierarchical camera sampling: Avoiding complex viewpoint selection optimization, it achieves sufficient multi-view coverage using spherical binning + uniform sampling + random truncation. This "good enough" engineering practicality allows multi-view densification to be seamlessly embedded into the standard 3DGS training loop without introducing intricate hyperparameters.

Limitations & Future Work

• Specular reflection upper limit is bounded by 1st-order SH: The authors acknowledge that 1st-order SH is inherently incapable of representing complex specular reflections captured by 3rd-order SH (e.g., perfect specular highlights, narrow specular lobes). Consequently, rendering quality drops significantly for scenes containing abundant specular surfaces (e.g., car showrooms, glass buildings).
• Peak training VRAM remains at the 3rd-order level: The first 3k iterations still require maintaining full 3rd-order SH (48 coefficients/Gaussian). The peak VRAM is comparable to standard 3DGS (appearing as 388 MB in the ablation table), which fails to lower the training bar despite the final model being lightweight—making on-device training on VRAM-constrained edge devices still challenging.
• Hierarchical sampling may miss critical narrow views: Multi-view pruning relies on the sampled camera subset to evaluate redundancy. If certain microstructures are only visible within an extremely narrow, unsampled view, they might be mistakenly pruned—this presents a higher risk in extremely sparse-view reconstruction scenes.
• Compression ratio of the quantization scheme is not deeply explored: Currently adopting the quantization scheme of Mobile-GS, matched coding strategies tailored to this compact 1st-order SH representation (such as codebook-based entropy coding or residual quantization) have not been explored—leaving this a direct path for further storage reduction.
• Dynamic scenes are not covered: The method currently supports only static scenes. Extending it to 4D GS (dynamic scenes) requires addressing temporal redundancy—the multi-view densification strategy could potentially be adapted as a multi-timestep consistency constraint, but this has not been explored in this paper.

Related Work & Insights

• vs Mobile-GS: Both being mobile GS schemes, Mobile-GS utilizes knowledge distillation + online MLP enhancement, whereas Flux-GS adopts MC energy aggregation + static baking enhancement. The core difference lies in Flux-GS shifting the high-frequency restoration calculation from the inference phase to a one-time operation during training, yielding several-fold training acceleration and zero inference overhead.
• vs C3DGS / LocoGS: These methods model view-dependent color or full Gaussian attributes using neural field architectures but require MLP forward passes during inference—Flux-GS's "pre-baking" strategy dominates directly in terms of FPS.
• vs EAGLES / PUP 3DGS: These pruning methods rely on transmittance or Hessian sensitivity analysis for importance ranking, still evaluating on a single-view basis. Flux-GS's multi-view alpha-weighted evaluation incorporates the view-consistency dimension, offering a more natural redundancy criterion—this logic can be transferred to other neural rendering methods requiring "pruning" (e.g., ray point filtering in NeRF-based methods).
• vs MVGS: MVGS is the first work to propose multi-view regularization for training GS, but its multi-view constraint operates at the loss level. Flux-GS utilizes multi-view information for densification and pruning decisions, offering a more low-level, direct structural multi-view guidance.

Rating

• Novelty: ⭐⭐⭐⭐ Utilizing directional moment estimation instead of distillation for SH compression is a novel idea, and the assessment perspective of multi-view alpha-weighted densification/pruning offers insights. However, the overall pipeline remains an incremental integration of existing components (MC sampling + MLP decoding + multi-view consistency).
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Includes main tables on 3 datasets, 5 groups of ablations, sensitivity analysis on K-value/camera counts/quantiles, per-scene results, user studies, and SH component visualizations. The experimental design is comprehensive and solid.
• Writing Quality: ⭐⭐⭐⭐ The methodology motivation is clear, with each component detailing "why it is done" and "how it improves over previous methods." The systematic difference analysis chapter comparing with Mobile-GS (Sec 4) is a great incremental benchmarking practice; the framework in Fig 3 and the densification comparisons in Fig 4 are intuitive and effective.
• Value: ⭐⭐⭐⭐⭐ Slashes training time from 80+ minutes to 11 minutes while maintaining competitive visual quality, with 2-5 MB storage and cross-platform WebGL availability—possessing extremely high practical value and directly lowering the barrier for 3DGS mobile deployment.

Related Papers

• [CVPR 2026] Energy-GS: Image Energy-guided Pose Alignment Gaussian Splatting with redesigned pose gradient flow
• [ICLR 2026] Mobile-GS: Real-time Gaussian Splatting for Mobile Devices
• [ICCV 2025] BANet: Bilateral Aggregation Network for Mobile Stereo Matching
• [CVPR 2026] Confidence-Guided Multi-Scale Aggregation for Sparse-View High-Resolution 3D Gaussian Splatting
• [ECCV 2026] Capacity-Controlled Multi-View Stylization of 3D Gaussian Splatting