Tetrahedral SDF Representation

• Tetrahedral SDF representation is a 3D modeling paradigm that uses tetrahedral meshes with signed distance fields to precisely capture surface details.
• It employs barycentric interpolation and marching tetrahedra for efficient volumetric rendering and analytic mesh extraction.
• Adaptive refinement and regularization losses ensure high-fidelity, real-time rendering with accurate, closed surfaces in 3D reconstruction.

A tetrahedral SDF representation is a 3D geometric modeling paradigm in which a signed distance field (SDF) is discretized and interpolated over the vertices of a tetrahedral mesh. This approach has become central to recent advances in 3D generative modeling, neural rendering, and high-fidelity mesh extraction, offering a blend of volumetric consistency, differentiability, and precise surface recovery.

1. Definition and Core Mathematical Structures

A tetrahedral SDF representation models a volumetric domain $\Omega \subset \mathbb{R}^3$ as a union of non-overlapping tetrahedra, each defined by four vertices $(v_a, v_b, v_c, v_d)$. At each vertex, a scalar SDF value $f_n \approx \mathrm{SDF}(v_n)$ is stored, where $\mathrm{SDF}(\cdot)$ denotes the signed distance to the target surface (negative inside, positive outside). For any point $x$ inside a tetrahedron $t$, barycentric linear interpolation yields

$$\mathrm{SDF}(x) = \sum_{i \in \{a, b, c, d\}} \lambda_i f_i,$$

where $\lambda_i$ are the barycentric coordinates of $x$ with respect to the tetrahedron’s vertices. This interpolation ensures that the SDF is continuous and piecewise-linear throughout the mesh.

Because the interpolant is affine within each tetrahedron, the SDF gradient $\nabla \mathrm{SDF}(x)$ is constant in each cell and can be precomputed from the vertex values. This piecewise-linear structure is essential for both efficient evaluation and analytic isosurface extraction (Gu et al., 2024, Cui et al., 26 Apr 2026, Oh et al., 20 Nov 2025).

2. Construction and Adaptive Refinement of Tetrahedral Grids

Tetrahedral meshes are generated either by structured subdivision (e.g., subdividing a regular grid into tetrahedra) or by adaptive approaches such as Delaunay tetrahedralization over a set of 3D points $(v_a, v_b, v_c, v_d)$0 covering the domain. Adaptive refinement is frequently employed to focus resolution near complex geometric features:

• Densification criteria include large circumradius, straddle of the zero level set, or high photometric error.
• Pruning operations remove tetrahedra or vertices that contribute minimally to the visible surface, using measures based on SDF magnitude or accumulated rendering contribution (Cui et al., 26 Apr 2026).

Some frameworks (e.g., TetraSDF) employ multi-resolution hash-grid encodings to allow for fine-grained positional encoding and feature aggregation across scales, aiding both expressivity and memory efficiency (Oh et al., 20 Nov 2025).

3. Volume Rendering and Differentiable Rasterization

Surface-based volumetric rendering leverages the spatially localized SDF field to efficiently approximate color integration along viewing rays. Each tetrahedron, treated as a volumetric primitive, contributes opacity according to the SDF’s variation across its entry and exit points on the ray:

$(v_a, v_b, v_c, v_d)$1

where $(v_a, v_b, v_c, v_d)$2 is the logistic CDF mapping controlled by steepness $(v_a, v_b, v_c, v_d)$3. The limit $(v_a, v_b, v_c, v_d)$4 recovers sharp interfaces identical to the mesh isosurface (Gu et al., 2024). Alpha compositing accumulates the contributions for color, depth, and normals. All operations—including barycentric interpolation, alpha computation, and per-pixel sorting—are made differentiable with custom CUDA kernels (Gu et al., 2024, Cui et al., 26 Apr 2026).

Rasterization-based frameworks partition the image plane into tiles (e.g., $(v_a, v_b, v_c, v_d)$5 pixels), cull tetrahedra not overlapping any tile, and conduct parallel compositing of the surviving fragments. This approach enables real-time rendering with frame rates of 15–25 FPS at $(v_a, v_b, v_c, v_d)$6 resolution, while maintaining differentiability for optimization (Gu et al., 2024).

4. Mesh Extraction and Analytic Isosurface Recovery

Mesh extraction from a tetrahedral SDF is performed via the Marching Tetrahedra algorithm. For each tetrahedron, zero-crossings on edges connecting vertices with opposite SDF signs are located by interpolation:

$(v_a, v_b, v_c, v_d)$7

A precomputed case table specifies the triangles to be generated, guaranteeing closed, watertight isosurfaces at $(v_a, v_b, v_c, v_d)$8 (Gu et al., 2024, Cui et al., 26 Apr 2026).

Advanced representations such as TetraSDF further enable analytic isosurface extraction by leveraging continuous piecewise affine (CPWA) structure. By composing a multi-resolution tetrahedral encoder with a ReLU MLP, the entire SDF becomes CPWA, and mesh extraction tracks both grid-induced and ReLU-induced linear region boundaries, achieving highly self-consistent meshes that match the learned SDF exactly (Oh et al., 20 Nov 2025).

5. Regularization and Optimization Objectives

To ensure SDF validity and geometric quality during optimization, several regularization losses are used:

• Eikonal loss: Penalizes deviation of SDF gradient norm from unity,

$(v_a, v_b, v_c, v_d)$9

where $f_n \approx \mathrm{SDF}(v_n)$0 is the per-tetrahedron SDF gradient.

• Normal consistency loss: Encourages alignment of adjacent tetrahedron or mesh vertex normals, suppressing spurious high-frequency “crinkles” (Gu et al., 2024).
• Mesh-to-field consistency and Laplacian curvature losses: Guide the zero-level set surface to be both smooth and consistent with field-derived geometry (Cui et al., 26 Apr 2026).
• Photometric and appearance losses: Enforce agreement between rendered and ground-truth color or diffusion-prior expectations in generative pipelines.

Optimization is typically performed with Adam, leveraging efficient first-order gradients propagated through the hash-grid, SDF network, and rendering/rasterization modules (Gu et al., 2024, Oh et al., 20 Nov 2025, Cui et al., 26 Apr 2026).

6. Implementation Performance and Quantitative Results

Recent implementations (e.g., Tetrahedron Splatting, SDFRaster, TetraSDF) demonstrate high efficiency and accuracy:

• Geometry stage convergence in 40–98 minutes per scene on contemporary GPUs.
• Real-time rendering at 15–25 FPS with $f_n \approx \mathrm{SDF}(v_n)$1 pixel grids (Gu et al., 2024).
• Quantitative metrics such as Chamfer distance (e.g., 0.68 mm on DTU (Cui et al., 26 Apr 2026)), geometry and appearance CLIP scores, and network-vs-mesh self-consistency (e.g., SSDF$f_n \approx \mathrm{SDF}(v_n)$2 with analytic extraction (Oh et al., 20 Nov 2025)).
• Meshes are typically hole-free and faithful to the field, outperforming grid-based methods in both precision and structural consistency.

7. Relations to Prior Work and Research Significance

Tetrahedral SDF techniques fuse strengths of several modeling paradigms:

• Unlike NeRF-style volumetric grids, they enable precise, analytic surface extraction and avoid the computational and topological limitations of regular grids.
• Compared to point-based splatting (e.g., 3DGS), tetrahedral SDFs support clean, closed surfaces and reliable mesh extraction, while retaining rasterization efficiency (Gu et al., 2024).
• CPWA-based frameworks such as TetraSDF introduce exact analytic meshing, eliminating sampling biases and enabling rigorous network-mesh self-consistency (Oh et al., 20 Nov 2025).
• Differentiable rasterization and “mesh-in-the-loop” losses (e.g., in SDFRaster) tightly couple appearance fitting and surface supervision for high-fidelity, multi-view-consistent 3D reconstruction (Cui et al., 26 Apr 2026).

The ability to represent, render, and extract accurate meshes in a single, differentiable pipeline has positioned tetrahedral SDFs as a principal representation for neural 3D generation, mesh reconstruction, and differentiable rendering research.