3D Gaussian Splatting & WRF-GS

• 3D Gaussian Splatting is a technique that models scenes with anisotropic Gaussian primitives, while WRF-GS extends this with frequency-aware and weighted loss components.
• WRF-GS employs wavelet decomposition to separate low- and high-frequency information, enabling targeted optimization for both global structure and fine details.
• Empirical results demonstrate that WRF-GS improves reconstruction fidelity, outperforming standard methods in metrics like SSIM, PSNR, and LPIPS.

3D Gaussian Splatting (WRF-GS) describes a family of methods for explicit 3D scene representation and rendering in which spatial structure, color, and sometimes physical or semantic fields are modeled as mixtures of anisotropic Gaussian primitives. The Weighted Residual Formulation for Gaussian Splatting (WRF-GS) specifically denotes the integration of wavelet domain analysis or weighted loss strategies into the Gaussian Splatting pipeline, promoting frequency-aware optimization and improved structure/detail fidelity. WRF-GS has been developed as both a comprehensive framework for scene reconstruction (Zhao et al., 16 Jul 2025) and as a conceptual extension for physically-aware loss weighting (Byrski et al., 31 Jan 2025), and is sometimes used as a generic abbreviation in survey literature for frequency-decoupled or physically weighted Gaussian methods (Bao et al., 2024).

1. Principles of 3D Gaussian Splatting and the WRF-GS Paradigm

In 3D Gaussian Splatting, a scene is represented by a set of anisotropic Gaussian primitives $$G = \{ (\mu_i,\,\Sigma_i,\,c_i,\,\alpha_i) \}_{i=1}^N$$, where each primitive has a center $\mu_i \in \mathbb R^3$, a covariance $\Sigma_i \in \mathbb R^{3\times3}$ (symmetric positive definite), a color or radiance parameter $c_i$, and an opacity $\alpha_i$ controlling its volumetric contribution. Rendering is achieved by projecting each Gaussian into screen space, producing elliptical 2D "splats"; accumulated radiance is then composited using front-to-back alpha blending (emission–absorption volume rendering) (Matias et al., 20 Oct 2025, Bao et al., 2024, Han et al., 5 Sep 2025).

Weighted Residual Formulation (WRF-GS) is an extension that introduces either 3D spatial frequency decomposition (wavelet domain) or per-primitive weighted objectives into this pipeline. The Wavelet-GS approach decomposes the input (point cloud, geometry, or initial Gaussian field) into low- and high-frequency bands using a 3D Discrete Wavelet Transform (DWT), which allows for distinct optimization strategies per band and explicit architectural separation of global structure and fine detail (Zhao et al., 16 Jul 2025).

2. Wavelet-Based Frequency Decoupling in WRF-GS

The Wavelet-based WRF-GS pipeline begins by applying a 3D DWT to the initial point set, yielding $P = P_{\text{low}} + P_{\text{high}}$ where $P_{\text{low}}$ (approximation coefficients) encode low-frequency, large-scale structures and $P_{\text{high}}$ (detail coefficients) encode local geometry and reflectance anomalies. Both components are independently voxelized and mapped to corresponding Gaussian primitives via branch-specific MLPs:

• $$G_{\text{low}}(x_l) = \{ \sigma_l, \mu_l, \alpha_l, c_l \}$$, capturing structure and smooth color fields.
• $$G_{\text{high}}(x_h) = \{ \sigma_h, \mu_h, \alpha_h, c'_h \}$$, refining detail and providing residual corrections (Zhao et al., 16 Jul 2025).

This decoupling allows the network to simultaneously optimize global consistency (over P_low) and local photorealism (over P_high) using targeted loss terms, including pixel-based, structural similarity, volumetric, and wavelet-domain (Laplacian-Wavelet) losses. A key component is a relighting module that operates on high-frequency Gaussians via SH-based networks, supervised by 2D image wavelet coefficients, to alleviate lighting artifacts and improve realism.

3. Rendering, Optimization Pipeline, and Loss Formulations

The rendering stage involves projection of 3D Gaussians onto the image plane and compositing as per classical or differentiable splatting:

• For each view, each Gaussian $\mu_i \in \mathbb R^3$0 is projected to pixel coordinates using camera intrinsics and extrinsics, with image-space covariance computed as $\mu_i \in \mathbb R^3$1 (Jacobian of the camera mapping).
• For each pixel, the color is composited as $\mu_i \in \mathbb R^3$2 where $\mu_i \in \mathbb R^3$3 are visible splats at $\mu_i \in \mathbb R^3$4 sorted by depth.

Optimization follows an alternating or joint schedule:

1. Forward: For a sampled set of rays or pixels, splat both low- and high-frequency Gaussians using the respective MLP branches and render predicted images.
2. Loss Computation: Aggregate per-pixel $\mu_i \in \mathbb R^3$5, $\mu_i \in \mathbb R^3$6, volumetric, Laplacian-Wavelet, and spherical-harmonic environment losses.
3. Backpropagate: Compute gradients w.r.t. all network, Gaussian, and relight parameters.
4. Adaptive Control: Split or prune Gaussians based on significance (opacity × volume × gradient magnitude), refine voxelization as needed.

The overarching loss in Wavelet-GS is typically

$\mu_i \in \mathbb R^3$7

where $\mu_i \in \mathbb R^3$8 is a weighted sum of $\mu_i \in \mathbb R^3$9 and $\Sigma_i \in \mathbb R^{3\times3}$0, $\Sigma_i \in \mathbb R^{3\times3}$1 promotes plausible SH environment terms derived from wavelet features, and $\Sigma_i \in \mathbb R^{3\times3}$2 enforces multi-scale consistency between rendered and reference images (Zhao et al., 16 Jul 2025).

4. Physical and Algorithmic Extensions: Weighted Loss and Ray-based WRF-GS

A variant of WRF-GS involves explicit weighted-residual learning in both rasterization-based and ray-tracing settings. For example, in RaySplats (Byrski et al., 31 Jan 2025), a per-Gaussian residual weighting scheme can be embedded into the loss function and ray solver:

$\Sigma_i \in \mathbb R^{3\times3}$3

with $\Sigma_i \in \mathbb R^{3\times3}$4 denoting the adaptive residual weight for Gaussian $\Sigma_i \in \mathbb R^{3\times3}$5. The ray–ellipsoid intersection step can be modified to increase the influence of high-residual Gaussians by inflating their effective support, or by biasing ray traversal ordering and early-termination thresholds using $\Sigma_i \in \mathbb R^{3\times3}$6 (Byrski et al., 31 Jan 2025).

In physical domains, such as wireless radiation field (RF) modeling, WRF-GS can generalize to complex-valued amplitudes with neural networks predicting amplitude and phase per Gaussian, incorporating multipath propagation and electromagnetic wave physics directly into the splatting and compositing process (Wen et al., 2024).

5. Quantitative Results and Empirical Findings

Experimental comparisons in (Zhao et al., 16 Jul 2025) demonstrate that the Wavelet-GS formulation (WRF-GS) yields state-of-the-art performance across scene reconstruction benchmarks, with consistent improvements over standard 3DGS and octree-based variants:

Method	SSIM ↑	PSNR ↑	LPIPS ↓
3DGS	0.840	27.19	0.313
Octree‐GS	0.828	26.82	0.321
Wavelet‐GS (WRF-GS)	0.853	28.34	0.274

Qualitative analysis indicates more complete extraction of global geometry (e.g., structure outlines in LiDAR/urban datasets) and sharper local effects (e.g., specular, shadow details) (Zhao et al., 16 Jul 2025). WRF-GS demonstrates robust convergence and improved resistance to high-frequency noise, especially in scenes with challenging lighting or texture patterns.

A plausible implication is that frequency-aware regularization and dual-branch architectures mitigate over-smoothing and facilitate higher-fidelity reconstruction, as corroborated by ablation results in (Zhao et al., 16 Jul 2025).

6. Applications, Extensions, and Future Challenges

WRF-GS and its derivatives extend naturally to a wide range of domains:

• Surface reconstruction: Via render-then-fuse pipelines, enabling watertight mesh extraction from splatted scenes (Ye et al., 2024).
• RF field modeling: Neural WRF-GS achieves substantial improvements in RF spectrum and channel state prediction, outperforming NeRF and conventional models by measurable dB gains (Wen et al., 2024).
• Large-scale visualization: Multi-GPU and distributed training regimes have enabled reconstructions with tens of millions of Gaussians interactively (Han et al., 5 Sep 2025).
• Hybrid and generative modeling: Integration with SDFs, triplanes, and diffusion models for feed-forward or conditional scene synthesis (see extensions surveyed in (Bao et al., 2024)).

Outstanding research challenges include further memory reduction (e.g., exploiting vector quantization and significance pruning), high-frequency detail recovery (potentially via multi-level wavelet decompositions), and extension to dynamic and non-Lambertian scenes (Zhao et al., 16 Jul 2025, Bao et al., 2024). Additionally, uncertainty quantification, adaptive control, and federated learning approaches are identified as promising future directions (Han et al., 5 Sep 2025).

---

References:

• "Wavelet-GS: 3D Gaussian Splatting with Wavelet Decomposition" (Zhao et al., 16 Jul 2025)
• "RaySplats: Ray Tracing based Gaussian Splatting" (Byrski et al., 31 Jan 2025)
• "Neural Representation for Wireless Radiation Field Reconstruction: A 3D Gaussian Splatting Approach" (Wen et al., 2024)
• "Toward Distributed 3D Gaussian Splatting for High-Resolution Isosurface Visualization" (Han et al., 5 Sep 2025)
• "From Volume Rendering to 3D Gaussian Splatting: Theory and Applications" (Matias et al., 20 Oct 2025)
• "3D Gaussian Splatting: Survey, Technologies, Challenges, and Opportunities" (Bao et al., 2024)