Geometry-Grounded Gaussian Splatting

• Geometry-grounded Gaussian splatting is a scene representation technique that embeds explicit geometric priors into Gaussian primitives for improved rendering and consistency.
• It leverages geometric features like normals, curvature, and surface descriptors to guide initialization and optimization, ensuring accurate surface alignment.
• This method achieves superior view synthesis, mesh extraction, and dynamic reconstructions by enforcing multiview consistency and physical plausibility.

Geometry-grounded Gaussian splatting refers to a class of approaches that imbue the optimization, initialization, or deformation of 2D/3D Gaussian primitives for scene representation with explicit geometric information, constraints, and priors. These methods enhance surface fidelity, rendering accuracy, and multi-view consistency by exploiting or estimating scene geometry—such as normals, curvature, SDFs, or multiview correspondences—at all stages of the pipeline. Geometry grounding contrasts with naïvely photometric losses or purely implicit deformation networks, which are insufficient for consistent or accurate surface modeling under varying viewpoints or sparse observations.

1. Fundamentals of Geometry-Grounded Gaussian Splatting

Geometry-grounded splatting models a scene as a collection of 3D anisotropic Gaussians, each parameterized by a mean $\mu_i \in \mathbb{R}^3$, covariance $\Sigma_i \in \mathbb{R}^{3\times 3}$ (often factored via principal axes and scales), color coefficients $c_i$, and opacity $\alpha_i$. Rendering is performed by projecting these Gaussians into camera space and compositing their contributions along camera rays using either volumetric integration or alpha-compositing.

Unlike standard splatting pipelines, geometry-grounded approaches use explicit geometric features—such as surface normals, curvature, SDF values, planar descriptors, or MVS-derived depth—to:

• Initialize Gaussian parameters so primitive placement, orientation, and aspect ratio match the underlying surface.
• Structurally regularize optimization and densification procedures (e.g., constraining duplicate placement to tangent spaces).
• Incorporate geometric consistency losses, including multiview depth consistency, normal alignment, or planarity.
• Inform deformation or blending fields when modeling dynamic scenes, ensuring local deformations respect scene structure.

These strategies yield a representation in which each Gaussian is "aware" of or "tied to" explicit 3D scene properties, improving both visual and geometric fidelity for synthesis, relighting, and mesh extraction tasks (Zhang et al., 25 Jan 2026, Lu et al., 2024, Li et al., 5 Sep 2025, Li et al., 2024).

2. Incorporation of Geometric Priors in Representation and Initialization

Geometry grounding begins at initialization. Several works construct geometry-aware Gaussians aligned to surfaces extracted from point clouds or MVS:

• Surface-aligned initialization: Gaussian means are placed directly on reconstructed surface vertices from SfM/MVS. Anisotropic covariances are configured so that the smallest axis aligns with the estimated surface normal and has subpixel thickness, while tangent axes are scaled by curvature or local geometric complexity (Li et al., 2024, Li et al., 5 Sep 2025, Wang et al., 1 Jul 2025).
• Curvature-guided initialization: GeoSplat computes principal curvature directions and values at each point, aligning tangential axes with principal directions and scaling the axes by the reciprocal of the curvatures to ensure local surface coverage without redundancy (Li et al., 5 Sep 2025).
• MLP-parameterized initialization: GDGS employs a learnable regressor to lift sparse 3D points into optimal Gaussian parameters, ensuring initial mean/covariance proximity to the true surface (Wang et al., 1 Jul 2025).

This geometry-informed placement prevents Gaussians from drifting off-surface in low-texture or ambiguous regions and leads to much faster and more stable convergence compared to random or color-driven initialization.

3. Geometry-Aware Optimization, Losses, and Regularization

Geometry-grounded splatting employs geometric regularizers throughout optimization:

• Normal and Plane Alignment: Gaussians are penalized for misalignment between their primary orientation and local surface normal, as well as for deviating from fitted planes in planar regions (Li et al., 2024, Zanjani et al., 2024, Li et al., 5 Sep 2025). Operations such as splitting and cloning are constrained to the tangent plane.
• Curvature- and Manifold-Driven Updates: Updates to positions and densification procedures project movement and placement offsets onto the tangent plane, or distribute new Gaussians based on local curvature, avoiding "floating" and "needle"-shaped artifacts (Li et al., 5 Sep 2025).
• Multiview and Feature-Based Consistency: When available, multiview stereo or learned feature consistency across views drives additional losses that enforce cross-view surface agreement, robust to failures in weakly textured areas (Kim et al., 16 Jun 2025, Wu et al., 29 Apr 2025).

For dynamic scenes, geometry-aware deformation models condition Gaussian translations and rotations on geometry descriptors extracted via 3D sparse convolutional networks and fused with auxiliary identity features, yielding local deformation fields coherent with the static scene geometry (Lu et al., 2024).

4. Volumetric and Field-Theoretic Formulations

Recent theoretical advances recast Gaussian primitives as volumetric stochastic solids, bringing rigor to surface extraction and rendering:

• Stochastic solid framework: Each Gaussian is interpreted as defining a probabilistic occupancy field, whose zero-level set provides an explicit geometric surface. Volume-rendering is performed via closed-form transmittance and attenuation coefficients derived from this field, enabling precise depth map extraction without floating-point artifacts or multi-view inconsistency as in prior pointwise alpha-accumulation schemes (Zhang et al., 25 Jan 2026, Jiang et al., 2024).
• Discretized SDF embedding: Relightable assets are produced by embedding a sampled SDF value into each Gaussian. This discretized SDF is mapped to opacity using a nonlinear transfer function and regularized by projection-based losses, aligning the reconstructed surface with the true geometry without requiring dense field networks or per-ray marching (Zhu et al., 21 Jul 2025).
• Planar and surfel geometry field models: Several methods explicitly enforce piecewise-planar structure and group Gaussians via planar descriptors built from image segmentation and surface normals, clustering them with probabilistic tree-based mixtures (Zanjani et al., 2024). Surfel-based geometry fields provide nearly exact, differentiable rasterization, crucial for reconstructing opaque surfaces with high curvature or discontinuities (Jiang et al., 2024).

5. Applications: Dynamic View Synthesis and Robust 3D Reconstruction

Geometry-grounded Gaussian splatting has demonstrated superior performance across multiple vision tasks:

• Dynamic view synthesis: Geometry-aware deformation models yield temporally coherent, photometrically faithful animation, sharply capturing locally intricate motions (e.g., articulated limbs, thin spines) and preserving canonical point cloud alignment over time. On D-NeRF and HyperNeRF dynamic datasets, PSNR and SSIM are improved versus implicit field baselines and prior dynamic Gaussian methods (Lu et al., 2024).
• Surface reconstruction and mesh extraction: Geometry-informed pipelines consistently achieve state-of-the-art on DTU and Tanks & Temples geometry evaluations, with lower Chamfer distances and higher mesh completeness compared to standard 3DGS, SuGaR, and 2DGS. SDF-embedded, quantile-densified, and surfel-splatted approaches yield more robust recovery of thin, complex, and specular structures (Zhang et al., 25 Jan 2026, Jiang et al., 2024, Zhu et al., 21 Jul 2025, Wu et al., 29 Apr 2025, Li et al., 2024).
• Relightable and generalizable assets: Discretized SDFs with deferred shading enable photorealistic material decomposition and real-time relighting with reduced memory compared to volumetric neural SDF baselines (Zhu et al., 21 Jul 2025). Generalizable pipelines such as G³Splat achieve state-of-the-art geometry recovery and relative pose estimation under pose-free self-supervision (Hosseinzadeh et al., 19 Dec 2025).

Key Quantitative Results (Scene and Task Examples)

Method	PSNR↑ / SSIM↑	Chamfer (mm)↓	Mesh F1↑	LPIPS↓
GeoSplat	36.37 / .976	—	—	.024
PlanarGS	—	—	—	—
Discretized SDF-GS	24.5 / .923	0.0107	—	.0762
Geometry-FieldSurf.	—	0.60	—	—
Sparse2DGS	—	1.13	—	—
EGGS	27.96 / .851	0.91	—	.192
G³Splat	—	0.244	—	—
Geometry-GroundedGS	—	0.47	0.60	—

Statistics as reported: DTU, Mip-NeRF360, BlendedMVS, ScanNet, and standard benchmarks (Zhang et al., 25 Jan 2026, Lu et al., 2024, Li et al., 5 Sep 2025, Zhu et al., 21 Jul 2025, Wu et al., 29 Apr 2025, Zhang et al., 2 Dec 2025, Hosseinzadeh et al., 19 Dec 2025, Jiang et al., 2024).

6. Extensions, Limitations, and Open Challenges

Geometry-grounded splatting is evolving rapidly, with several directions highlighted in recent work:

• Extension to higher-order geometric descriptors (e.g., torsion, surfaces with singularities) to better address non-manifold or filamentary structures (Li et al., 5 Sep 2025).
• Improved robustness of geometric prior estimation in low-SNR or occluded regions, and integration with learned semantic or foundation model cues (e.g., CLIP-guided planar proposals) (Zanjani et al., 2024).
• Efficient parallelization of curvature, normal, or feature estimation, as large-scale scene modeling challenges memory and computation (Li et al., 5 Sep 2025, Li et al., 2024).
• Unification with advanced relighting, material decomposition, and generalizable pipelines for real-world deployment and AR/VR integration (Zhu et al., 21 Jul 2025, Hosseinzadeh et al., 19 Dec 2025).
• Handling failure cases where geometry assumptions (local smoothness, planarity) break down, especially in scenes with extreme topology or reflectance.

The stochastic-solid theoretical foundation and diverse algorithmic instantiations collectively establish geometry-grounded Gaussian splatting as the new state of the art for efficient, accurate, and physically interpretable scene representation.

---

Notable References:

• 3D Geometry-aware Deformable Gaussian Splatting for Dynamic View Synthesis (Lu et al., 2024)
• Geometry Field Splatting with Gaussian Surfels (Jiang et al., 2024)
• Planar Gaussian Splatting (Zanjani et al., 2024)
• GeoSplat: A Deep Dive into Geometry-Constrained Gaussian Splatting (Li et al., 5 Sep 2025)
• Gaussian Splatting with Discretized SDF for Relightable Assets (Zhu et al., 21 Jul 2025)
• Multiview Geometric Regularization of Gaussian Splatting for Accurate Radiance Fields (Kim et al., 16 Jun 2025)
• GeoGaussian: Geometry-aware Gaussian Splatting for Scene Rendering (Li et al., 2024)
• Geometry-Grounded Gaussian Splatting (Zhang et al., 25 Jan 2026)
• Sparse2DGS: Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views (Wu et al., 29 Apr 2025)
• EGGS: Exchangeable 2D/3D Gaussian Splatting for Geometry-Appearance Balanced Novel View Synthesis (Zhang et al., 2 Dec 2025)
• G3Splat: Geometrically Consistent Generalizable Gaussian Splatting (Hosseinzadeh et al., 19 Dec 2025)
• MVG-Splatting: Multi-View Guided Gaussian Splatting (Li et al., 2024)
• GDGS: 3D Gaussian Splatting Via Geometry-Guided Initialization And Dynamic Density Control (Wang et al., 1 Jul 2025)