3D Surface Splatting (3DSS): Techniques & Trends

• 3D Surface Splatting (3DSS) is a method for real-time novel-view synthesis and high-fidelity surface extraction using explicit surfels and differentiable rendering.
• It employs anisotropic Gaussian kernels and joint photometric, geometric, and regularization optimizations to capture fine details and maintain global surface coherence.
• Advanced variants integrate implicit SDFs and volumetric fields to handle complex materials, semi-transparency, and large-scale scenes for enhanced rendering performance.

3D Surface Splatting (3DSS) is a family of advanced surface-based scene representations and differentiable rendering techniques that leverage explicit, spatially localized 3D kernels—predominantly anisotropic Gaussians or their generalizations—for real-time, photorealistic novel-view synthesis and high-fidelity surface mesh extraction. By jointly optimizing geometric, photometric, and regularizing objectives, 3DSS methods achieve a unique synergy between the photometric detail of rasterized primitives and the global coherence of surface constraints, enabling direct and flexible pipelines for surface reconstruction, material estimation, and inverse rendering in both small-scale and large-scale scenes. The approach unifies developments from classical point-based rendering, differentiable geometry processing, and hybrid radiance field methods, addressing critical limitations of mesh- or voxel-based models and neural implicit fields.

1. Mathematical Foundations of 3D Surface Splatting

3DSS represents a scene as a collection of N explicit surface samples (“surfels”), each modeled as an anisotropic 3D Gaussian kernel parameterized by its center $\mu_i \in \mathbb{R}^3$, covariance matrix $\Sigma_i \in \mathbb{R}^{3\times 3}$, color parameters (often as spherical harmonics coefficients) $c_i$, and opacity $\alpha_i$ or closely related density parameter. The forward rendering process blends the contributions of splats projected onto the image plane via front-to-back compositing:

$$G_i(x) = \alpha_i~\exp \left( -\tfrac{1}{2}(x - \mu_i)^\top\,\Sigma_i^{-1}(x - \mu_i) \right)$$

$$C(p) = \sum_{i} T_i \alpha_i c_i, ~~~~ T_i = \prod_{j < i} (1 - \alpha_j)$$

The anisotropy and orientation of $\Sigma_i$ enable tight adherence to local surface structure by “flattened” Gaussians (i.e., strong compression along the surface normal). The coverage-based compositing model in recent 3DSS (e.g., (Younes et al., 7 May 2026)) defines per-layer opacities $\alpha_k$ via normalized EWA kernel sums, ensuring unbiased anti-aliased silhouettes and differentiable surface separation.

Beyond raw surfel properties, 3DSS often incorporates explicit geometric attributes (normals $n_i$, depths $d_i$), local material parameters (albedo, roughness), and, in high-end pipelines, microfacet BRDF shading under HDR environment illumination. Layered compositing allows multiple surface intervals per pixel, supporting partial coverage and multi-modal transmittance essential for semi-transparent and complex layered scenes (Xu et al., 13 Nov 2025).

2. Optimization Protocols for Surface Fidelity

The optimization of 3DSS models is conducted end-to-end using photometric, geometric, and regularization losses. The central objective is typically an $\Sigma_i \in \mathbb{R}^{3\times 3}$0/D-SSIM photometric loss between rendered and observed images, mediated via the kernel-based rasterizer:

$\Sigma_i \in \mathbb{R}^{3\times 3}$1

To guarantee the correct spatial localization of surfels, approaches such as scale regularization and principal-axis flattening (Chen et al., 2023) reduce the smallest eigenvalue of $\Sigma_i \in \mathbb{R}^{3\times 3}$2, aligning surfel centers and normal directions with the true surface, reinforced by normal-consistency or normal-prior losses:

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

Global geometric coherence is enforced through SDF- or MVS-guided losses, Eikonal constraints (for enforcing unit gradient norm in SDFs), surface-consistency (pushing surfel centers onto $\Sigma_i \in \mathbb{R}^{3\times 3}$4), depth-distortion, and multi-view re-projection errors (Lyu et al., 2024, Gao et al., 21 Jul 2025, Gu et al., 18 Nov 2025). For scenes lacking reliable photometric cues (texture-less areas), monocular normal priors from pretrained backbone models supplement supervision (Wang et al., 2024, Shen et al., 2024).

Pipelines for mesh extraction typically render a depth map from the optimized surfel cloud and fuse these via volumetric TSDF techniques, often with subsequent Poisson surface reconstruction to ensure watertightness (Zhao et al., 2024, Chen et al., 21 Jun 2025).

3. Hybridization with Implicit and Volumetric Fields

Recent 3DSS research fuses explicit Gaussian splatting with implicit signed distance fields (SDFs), obtaining complementary benefits: (i) explicit splats concentrate high-frequency detail and sparse surface cues, (ii) implicit SDFs provide continuous, watertight geometry and dense off-surface regularization (Lyu et al., 2024, Chen et al., 2023).

3DGSR (Lyu et al., 2024) introduces a differentiable SDF-to-opacity mapping $\Sigma_i \in \mathbb{R}^{3\times 3}$5 (using a bell-shaped logistic derivative), connecting SDF distance to surfel opacity. The photometric loss jointly updates both SDF and Gaussian parameters; volume-rendered geometry attributes (depths and normals) from SDF and Gaussian branches are enforced to match, regularizing the SDF where splats are absent. The SDF is ultimately extracted via marching cubes.

SurfaceSplat (Gao et al., 21 Jul 2025) alternates two-way hybrid optimization: SDFs give coarse global geometry for initializing and regularizing 3DGS splats, while high-fidelity images rendered from optimized 3DGS serve as additional views for SDF refinement. Cyclic SDF↔3DGS alternation closes the loop, yielding state-of-the-art performance under sparse input.

MGSR (Zhou et al., 7 Mar 2025) operationalizes dual 2D-GS and 3D-GS branches, in which 2D flattening advances geometry fidelity while 3D-GS leverages geometry-guided, physically-motivated illumination decomposition. Alternating optimization, with mutual supervision via pseudo-ground-truth "transmitted" images and depth, yields both high-fidelity surface meshes and realistic NVS under diverse light conditions.

4. Adaptations for Sparse, Large-Scale, and Challenging Scenes

To address the ill-posedness of sparse-view 3D reconstruction, 3DSS methods employ:

• Geometry-prioritized initialization: Dense point clouds from learning-based MVS (e.g., CLMVSNet) seed surfel locations, normals, and scale (Wu et al., 29 Apr 2025, Gu et al., 18 Nov 2025).
• Stereo and pseudo-feature alignment: In SparseSurf (Gu et al., 18 Nov 2025), stereo geometry-texture alignment and pseudo-view feature consistency (including temporal or interpolated “virtual” views) bridge geometry and appearance across few-shot scenarios, mitigating overfitting and geometric drift from flattened (highly anisotropic) Gaussians.
• Consolidated solidness: SolidGS (Shen et al., 2024) replaces standard Gaussian tails with a solid kernel ($\Sigma_i \in \mathbb{R}^{3\times 3}$6 in the exponential), enforcing opacity locality and enabling robust geometry from as few as three views. Geometric regularization on interpolated views and monocular normal estimation further suppress ambiguity.
• Regularization in large-scale outdoor scenes: Adaptive binary partitioning and per-cell refinement localize spline support in aerial or urban contexts (Chen et al., 21 Jun 2025). Appearance decoupling (via learned correction maps) and transient-object masking (predicting soft masks for moving entities) disentangle lighting and dynamic effects, ensuring high-fidelity mesh extraction and robust large-scale surface reconstruction.

Coarse-to-fine reconstruction, dynamic partitioning, and local-to-global mesh merging strategies collectively enable scalable pipelines over kilometer-scale scenes at tractable computation and memory.

5. Extension to Complex Materials, Semi-Transparency, and High-Frequency Surfaces

Surface splatting methods have been extended to address complex reflectance, transparency, and fine-scale surface detail:

• View-dependent opacity and reflectance: VoD-3DGS (Nowak et al., 29 Jan 2025) augments each splat’s opacity with a symmetric matrix $\Sigma_i \in \mathbb{R}^{3\times 3}$7 controlling view-dependent attenuation as a quadratic function of viewing vector, allowing highlight/shadow modulation without deep networks. This approach supports real-time rendering with modest memory overhead and state-of-the-art metrics on large-scale relighting benchmarks.
• Semi-transparent and multi-layered surfaces: TSPE-GS (Xu et al., 13 Nov 2025) samples transmittance space rather than depth, allowing for multi-modal depth distributions on each pixel (multiple peaks for internal and external surfaces). Progressive TSDF fusion integrates multiple isosurfaces per view, producing accurate reconstructions for both opaque and semi-transparent objects without additional deep learning modules.
• High-frequency surface texture: 3D Gabor Splatting (Watanabe et al., 15 Apr 2025) enhances conventional Gaussian splats by multiplying with learnable sinusoidal (Gabor) kernels. Each primitive can represent oscillatory, high-frequency signals (e.g., stripes, textile patterns) within a single kernel, greatly reducing the primitive count necessary for fine detail. Results show unequivocal improvements in SSIM and PSNR for textured objects compared to baseline 2DGS at equal kernel budgets.

These variants confirm the malleability of 3DSS to a broad class of non-Lambertian and non-opaque materials.

6. Inverse Rendering, Anti-Aliasing, and Mesh Integration

3D Surface Splatting has been extended to physically-based inverse rendering, enabling joint recovery of geometry, spatially-varying BRDFs, HDR environment illumination, and material attributes (Younes et al., 7 May 2026). The core compositing model (interval merging of per-surfels’ support in view-space) yields anti-aliased silhouettes and informative visibility gradients for refinement and optimization. Microfacet BRDF shading (split-sum IBL) is computed at each surfel and composited at the pixel level, allowing direct relighting from multi-view imagery.

Adaptive refinement, driven by the magnitude of per-surfel screen-space positional gradients, increases sample density in under-resolved or high-variation regions, while zero-gradient or isolated surfels are pruned for efficiency. Mesh conversion is native: the surfel cloud (positions and normals) is fed to screened Poisson surface reconstruction or depth-fused via TSDFs, directly producing watertight triangle meshes suitable for conventional simulation, editing, or further rendering.

7. Comparative Performance, Limitations, and Future Directions

Empirical results across synthetic, controlled, and large-scale real datasets consistently demonstrate that 3DSS approaches achieve or surpass the geometric, photometric, and efficiency benchmarks set by mesh-based, voxel-based, and pure neural-implicit methods (Wang et al., 2024, Lyu et al., 2024, Gao et al., 21 Jul 2025, Chen et al., 21 Jun 2025). Typical metrics include Chamfer Distance, F-score, PSNR, SSIM, and LPIPS.

Strengths include:

• Robustness and accuracy in sparse-view settings where NeRF-style or MVS-based models degrade,
• High rendering speed (30–900 FPS depending on kernel count and resolution),
• Direct mesh extraction with minimal post-processing,
• Flexibility for material, transparency, and hybrid model integration.

Limitations remain in extremely under-constrained regions (e.g., single-view-only surfaces), unbounded scenes (for some pipelines), and highly complex dynamic deformation, though releases such as Dynamic Generalized Exponential Splatting (Zhao et al., 2024) and TranSplat (Kim et al., 11 Feb 2025) offer promising innovations. Potential areas for future work include integrating scene-wide connectivity priors, adapting explicit kernel functions for spatially adaptive or learned non-Gaussian bases, and leveraging hybrid surface-volume models for further robustness in unconstrained real-world data.

In summary, 3DSS provides a highly scalable, differentiable, and topologically flexible framework for achieving state-of-the-art results across surface reconstruction, novel-view synthesis, and physically-based inverse rendering. Its explicit-primitive foundation, versatility with geometric and photometric regularization, and capacity for hybridization position it as a cornerstone paradigm in modern computer vision and graphics research (Younes et al., 7 May 2026, Lyu et al., 2024, Gao et al., 21 Jul 2025, Chen et al., 2023, Shen et al., 2024, Wang et al., 2024).