Gaussian Splatting with Discretized SDF

• Gaussian splatting with discretized SDF is a hybrid 3D representation that couples explicit Gaussian primitives with structured SDF regularization for enhanced geometric accuracy.
• It employs SDF-derived opacity transfers and projection consistency to efficiently guide primitive placement, avoiding computationally expensive ray marching.
• This approach enables rapid novel-view synthesis, robust scene reconstruction, and scalable real-time mapping across diverse visual computing applications.

Gaussian Splatting with Discretized Signed Distance Fields (SDF)

Gaussian splatting with discretized signed distance fields (SDF) constitutes a hybrid representation that tightly couples the efficient, explicit rasterization of 3D Gaussian primitives with the global geometric coherence of SDF-based implicit surfaces. Recent advances leverage this synergy for high-fidelity 3D reconstruction, rapid novel-view synthesis, and robust geometry estimation across a range of visual computing domains, including scene reconstruction, robot mapping, and photorealistic relighting. Central to these techniques is the incorporation of discrete samples of an SDF—either per-Gaussian or on a voxel/octree grid—that regularize the placement, opacity, and pruning of splatting primitives and enable mesh extraction or global surface denoising. Distinct from pure NeRF/Neus-style volume rendering, these approaches avoid expensive ray marching on dense 3D grids, instead exploiting the structuring power of signed distance supervision to enhance the geometric consistency of point-based or grid-based explicit primitive renderers.

1. Mathematical Foundations and Hybrid Representations

A discretized SDF within the Gaussian splatting (GS) paradigm refers to a surface representation defined either by explicit samples of a signed distance function—for example, one value per Gaussian primitive, a dense voxel grid, or a sparse multiresolution hash grid. In the hybrid GS+SDF framework, the continuous SDF maps $f_\text{sdf}:\mathbb{R}^3\to\mathbb{R}$ whose zero-level set $\{\mathbf{x}\mid f_\text{sdf}(\mathbf{x})=0\}$ defines the surface. However, unlike volumetric implicit representations, discretization is leveraged for computational efficiency and tight primitive-surface coupling.

Examples of discretization modalities include:

• Per-Gaussian attribute: Each explicit 3D Gaussian primitive $g_i$ stores $s_i=f_\text{sdf}(\mu_i)$, the SDF evaluated at its center $\mu_i$ (Zhu et al., 21 Jul 2025).
• Hash grid / sparse grid: $f_\text{sdf}$ is represented as an MLP built atop a spatially sparse, multiresolution hash grid (Yu et al., 2024, Li et al., 2024). Trilinear (or multilinear) interpolation grants continuous evaluation everywhere, but storage and updates remain locally sparse.
• Dense voxel/octree: A regular voxel grid or adaptively subdivided octree $V^{\text{sdf}}$ encodes $f_\text{sdf}$ at specified spatial resolution, typically used for mesh extraction or geometric regularization (Gao et al., 21 Jul 2025, Li et al., 2024).

Gaussian primitives are parameterized by $(\mu_i,\,\Sigma_i,\,c_i,\,\alpha_i)$ for center, (anisotropic) covariance, appearance, and (possibly SDF-derived) opacity. Their explicit nature enables hardware-accelerated rasterization and real-time splatting, but without additional supervision, distribution and participation in forming surface geometry can be ill-defined in sparse or textureless regions (Xu et al., 2024, Xiang et al., 2024). The SDF, in continuous or discretized form, provides this absent global structure and regularizes both placement and blending of primitives.

2. SDF-to-Opacity Transfers and Surface Regularization

A key architectural feature in GS+SDF hybrids is the use of the SDF value (often per Gaussian) to control opacity via a nonlinear transfer function, typically modeled after the derivative of the sigmoid/logistic function. This transformation enforces that Gaussians with centers near the surface (i.e., $s_i\approx 0$) become highly opaque, while those far away vanish:

$$o_i = T_\gamma(s_i) = 4 e^{-\gamma s_i} [1+e^{-\gamma s_i}]^{-2}$$

with $\gamma$ a global or learnable sharpness parameter [(Zhu et al., 21 Jul 2025), 3DGSR (Lyu et al., 2024)].

This transfer admits efficient, non-iterative splatting: compositing each Gaussian according to its SDF-induced opacity yields sharp, shell-conforming surfaces without resorting to volumetric ray marching or expensive density integrations. Gaussian center positions can then be projected onto the SDF's zero-level set—$\mu_i^{\text{proj}}=\mu_i-s_i n_i$ with $n_i$ the per-Gaussian surface normal—enabling both geometric regularization (via projection-based consistency losses) and direct mesh extraction (Zhu et al., 21 Jul 2025).

Distinct from continuous SDF frameworks, where Eikonal and curvature losses regularize the global field, in discretized SDF-GS methods, alternative constraints such as depth/alignment errors between projected and rendered depths, or orientation consistency between SDF and Gaussian-inferred normals, are adopted to avoid needing field-wide gradients (Zhu et al., 21 Jul 2025, Xu et al., 2024). These constraints, combined with opacity control, mitigate floaters, enforce surface manifoldness, and yield globally coherent and sharp reconstructions.

3. Algorithmic Pipelines and Differentiable Splatting

Typical pipelines for GS with discretized SDF employ the following stages:

1. Initialization: Gaussians placed using heuristics (random scatter, mesh vertex sampling, COLMAP output, or via Marching Cubes on an SDF grid) (Xu et al., 2024, Liu et al., 13 Mar 2025, Gao et al., 21 Jul 2025).
2. Warm-up: Optimize only Gaussian positions and attributes for initial color consistency under splatting losses (Xu et al., 2024, Zhu et al., 2024).
3. Joint optimization: Alternating or simultaneous optimization of Gaussian parameters and SDF field under combined photometric, geometric, and regularization objectives. Losses include splatting-based color consistency, SDF alignment (projection losses, Eikonal, orientation, curvature), normal/depth consistency, and opacity entropy (for pruning semi-transparent non-surface Gaussians) (Xu et al., 2024, Li et al., 2024, Yu et al., 2024, Zhu et al., 21 Jul 2025).
4. Densification/pruning: Adaptive splitting or removal of Gaussians based on SDF-derived scores (e.g., $\eta = \exp(-S^2 / (\lambda_\sigma \sigma^2))$) and geometric confidence (Xiang et al., 2024, Li et al., 2024).
5. Differentiable rendering: Per-pixel compositing using alpha-blending of 2D-projected Gaussians, with SDF-derived opacity; often combined with deferred shading for physical relighting (Xu et al., 2024, Zhu et al., 21 Jul 2025, Zhu et al., 2024).

Mesh extraction is performed by standard Marching Cubes (or Tetrahedra) on the SDF field, using only a test-time grid, never requiring storage of a dense grid during learning [(Zhang et al., 2024), 3DGSR (Lyu et al., 2024)].

A summary of pipeline stages appearing in representative works is given below.

Stage	Description	Example References
Initialization	Scatter/mesh-based or grid-based Gaussian placement	(Xu et al., 2024, Liu et al., 13 Mar 2025)
SDF-to-opacity transfer	Nonlinear map from SDF sample to Gaussian opacity	(Zhu et al., 21 Jul 2025, Lyu et al., 2024)
Projection consistency	Depth/normal losses aligning SDF surfaces and GS projections	(Zhu et al., 21 Jul 2025, Xu et al., 2024)
Densification/pruning	Adaptive insertion/pruning of Gaussians using SDF/local scores	(Xiang et al., 2024, Li et al., 2024)
Mesh extraction	Marching Cubes/Tetrahedra on SDF at test time	(Zhang et al., 2024, Gao et al., 21 Jul 2025)

4. Applications and Empirical Performance

GS with discretized SDF achieves high-fidelity geometry and photorealistic rendering while maintaining real-time or near real-time performance, outperforming traditional voxel fusion and NeRF-type grids on efficiency and reconstruction quality.

Key areas of application and benchmark performance include:

• Surface reconstruction and relightable asset creation: Accurate, hole-free meshes suitable for BRDF-based rendering, with improved relighting PSNR and normal error compared to previous GS or hybrid methods (PSNR ≈ 24.5 dB, normal MAE ≈ 6.5° on Glossy Blender; surface-CD ≈ 0.0107) (Zhu et al., 21 Jul 2025).
• Robotic mapping and SLAM: SLAM pipelines leveraging point-based SDF+GS fusion demonstrate superior photometric/geometric metrics, improved odometry estimation, and memory efficiency (F-score↑2–4%; ARTE↓0.68% vs 1.00% on 5 km Oxford Spires; mapping scalability at reduced memory cost) (Pan et al., 9 Feb 2025, Liu et al., 13 Mar 2025, Peng et al., 15 Sep 2025).
• View synthesis and scene geometry under challenging conditions: Robustness under strong lighting (OO3D-SL) and semi-transparency (α‐NeuS), with reduced training/inference times (40 min – 1.3 h/object, interactive inference) relative to previous SDF+GS methods (Xu et al., 2024).
• Dynamic and large-scale scenes: GSDF hybrids supporting dynamic object modeling (trajectory bases, temporal SDFs), monocular/2D prior guidance, and scalable, hierarchical representations (Li et al., 2024, Tourani et al., 15 Oct 2025).

Empirically, discretized SDF-GS methods achieve state-of-the-art results on standard datasets, such as DTU, MobileBrick, OmniObjects3D-d, and ScanNet, with significant improvements in Chamfer distance, F-score, and perceptual rendering metrics compared to both pure GS and traditional implicit/voxel SDF baselines (Xu et al., 2024, Gao et al., 21 Jul 2025, Xiang et al., 2024, Yu et al., 2024).

5. Implementation Considerations and Computational Aspects

The GS+discretized SDF approach is designed for maximal computational efficiency with minimal memory overhead. Major implementation aspects include:

• Memory footprint: Absence of dense SDF grids during training (SDF only evaluated per-Gaussian or at test-time on a mesh-extraction grid) reduces total GPU memory usage to ≈4 GB for moderate scenes (Zhu et al., 21 Jul 2025). Some hybrid methods using TensoSDF/factorized grids require up to 20–22 GB (Zhu et al., 2024).
• Speed and scalability: Training time per object is reduced to 40 minutes–1.3 hours on flagship GPUs, a marked improvement over 2–16 hours required for prior SDF+GS hybrids (Xu et al., 2024, Zhu et al., 21 Jul 2025). Real-time (≥100 fps) inference is routinely achieved on 640×480 frames (Zhu et al., 21 Jul 2025). For SLAM, full-scene mapping on Azure Kinect runs at 150–200 fps (Peng et al., 15 Sep 2025).
• Hardware acceleration: CUDA-accelerated 2D rasterization, early-z rejection, depth sorting, and tile-based compositing are utilized to maximize throughput and minimize redundant computation (Xu et al., 2024, Zhu et al., 21 Jul 2025).
• Adaptive resource utilization: Hierarchical/voxel/octree-based discretization and dynamic pruning/densification mitigate both memory waste and geometric artifacts, adapting to local scene complexity (Gao et al., 21 Jul 2025, Li et al., 2024).

6. Limitations, Open Problems, and Future Directions

Despite empirical successes, the discretized SDF-GS paradigm presents several challenges and active research areas:

• Density and initialization dependency: Surface completeness and global smoothness are contingent on sufficient initialization density and coverage of Gaussian primitives. Extreme sparse-view inputs or pathological reflective phenomena can induce misplacement or coverage gaps (Xu et al., 2024, Xiang et al., 2024).
• Semi-transparent and multi-material surfaces: Directly modeling truly multi-layer or volumetrically complex materials is non-trivial in a single-layer Gaussian splatting regime; extensions with BRDF coupling or volumetric SDF integration are under exploration (Xu et al., 2024, Li et al., 2024).
• Boundary, occlusion, and global regularization: Discrete SDF samples and projection-based regularizers provide weaker guarantees than fully continuous Eikonal or curvature losses; further improvement in edge fidelity and manifold enforcement is an open topic (Zhu et al., 21 Jul 2025, Gao et al., 21 Jul 2025).
• Scalability and dynamic scenes: Hierarchical/multiresolution discretizations, adaptive densification, and temporal SDFs open promising directions for large-scale and dynamic scene reconstruction (Xu et al., 2024, Tourani et al., 15 Oct 2025).

Future work is directed toward adaptive and hierarchical Gaussian allocation, explicit mesh/material generation, dynamic scene modeling, and generalization beyond bounded indoor scenes—potentially leveraging hybrid SDF/primal-to-dual differentiable pipelines for unbounded and heterogeneous environments (Xu et al., 2024, Tourani et al., 15 Oct 2025).

7. Representative Methods and Comparative Summary

The spectrum of discretized SDF-GS architectures encompasses a variety of design axes: where, when, and how SDF supervision is applied; the nature of discretization (per-Gaussian, hash grid, voxel grid, or octree); mutual regularization between the explicit and implicit components; and application domains. Table summaries highlight several representative approaches.

Method	SDF Discretization	SDF-to-Opacity	Surface Regularization	Rendering/Use Case	Reference
GSurf	Continuous SDF @ Gaussian	L_pos, L_off, L_ori, etc.	Joint SDF/GS supervision	Real-time photo+mesh reconstruction	(Xu et al., 2024)
GS-SDF	Multires. hash grid	SDF-based Gaussian init	Shape/disk sampling losses	SLAM/LiDAR, robust geometry + image synthesis	(Liu et al., 13 Mar 2025)
3DGSR	Multi-res hash grid SDF	$\Phi_\beta$ (deriv. sig)	Eikonal; Gaussian-rooted	RGB (real/synth.), detailed meshes	(Lyu et al., 2024)
GaussianRoom	SDF MLP + voxel grid	SDF for densification	Eikonal, edge, normal priors	Indoor scene fidelity/completeness	(Xiang et al., 2024)
PINGS	Per-point SDF in hash voxel	Opacity from SDF/kNN	SLAM: chmfr, ARTE	LiDAR+camera SLAM, robot navigation	(Pan et al., 9 Feb 2025)
SurfaceSplat	Voxel grid SDF	Iso-anchored GS inits	L_iso, TV, Eikonal	Meshes from image, novel views	(Gao et al., 21 Jul 2025)
GS-Octree	SH SDF on octree grid	SDF guides/prunes GS	Eikonal, Hessian, TV	Strong-light recon, high real-time perf.	(Li et al., 2024)
GS-ROR$^2$	Factorized SDF (TensoSDF)	SDF mutual supervision	Mutual depth/normal, pruning	Relightable asset, specular object	(Zhu et al., 2024)
SplatSDF	3DGS-augmented SDF (train)	3DGS embedding fusion	Depth/normal consistency	High fidelity SDF with GS acceleration	(Li et al., 2024)
MonoGSDF	Multi-res hash grid	$\exp(-\beta\,\bar s_i^2)$	SF priors, multi-res	Monocular, fast, watertight meshes	(Li et al., 2024)
UGSDF	Multi-res grid MLP	Densify/prune by SDF	SDF, photo, eikonal	Dynamic/urban scenes, scene edits	(Tourani et al., 15 Oct 2025)

Taken together, these works establish the hybrid discretized SDF-Gaussian splatting paradigm as a leading approach for rapid, high-fidelity, geometrically consistent 3D vision and graphics, balancing explicit, efficient rasterization with the structure and denoising capabilities of implicit fields.