TetraCoder: Adaptive Tetrahedral Encoding for Neural Fields

Daniele Della Pietra, Antonella Rech, Nicola Conci, Nicola GarauarXiv 2026 · arXiv

TL;DR

TetraCoder is a novel spatial encoding for neural fields that uses a single-level, adaptive tetrahedral grid to bridge the gap between implicit (SDF) and explicit (mesh) geometry. By ensuring the representation is Continuous Piecewise Affine (CPWA) by design, it enables exact, lightning-fast analytic mesh extraction and localized scene editing without the artifacts or "spooky action at a distance" common in hash-based methods.

What Problem Does This Paper Address?

In the world of 3D computer vision and graphics, we often juggle two types of geometry representations:

  1. Implicit representations (like Signed Distance Fields or SDFs) are great for optimization and spatial queries but slow to turn into a surface for rendering or physics.
  2. Explicit representations (like meshes) are fast to render but harder to learn directly from raw data like images.

Current state-of-the-art methods, such as Instant-NGP, use multiresolution hash grids to capture fine details efficiently. However, they have two major drawbacks:

  • Hash Collisions: Updating a feature in one part of the scene might accidentally change a completely unrelated part because they share the same hash bucket.
  • Non-Affine Interpolation: Their trilinear interpolation doesn't produce a "piecewise-affine" field, making it mathematically messy and slow to extract an exact, high-quality mesh.

The authors of TetraCoder set out to create a representation that is as efficient as hash grids but provides a direct, exact link between the learned neural field and a physical mesh.

Methods

The core of TetraCoder is a single-level, hash-free adaptive tetrahedral grid.

  • Adaptive Grid: Instead of using multiple fixed-resolution grids, TetraCoder uses one grid that adapts during training. It uses densification (adding vertices where gradients are high), pruning (removing useless vertices), and movement (optimizing vertex positions) to concentrate detail exactly where the surface is.
  • CPWA by Design: By using barycentric interpolation within each tetrahedron and pairing it with ReLU-based neural decoders, the entire field becomes Continuous Piecewise Affine (CPWA). This means that within any small region, the geometry is defined by simple linear equations.
  • Analytic Mesh Retrieval: Because the field is CPWA, the "zero-level set" (the actual surface) can be calculated exactly as a collection of planar polygons. There’s no need for the approximation errors inherent in standard algorithms like Marching Cubes.
  • Tetrahedral Ray Tracing: To render the scene, the authors implement an efficient "adjacency walk" to find which tetrahedron a ray is currently passing through, significantly speeding up volume rendering.
Comparison of Hash Encoder, TetraSDF, and the proposed TetraCoder.

Datasets

The authors evaluated TetraCoder across several challenging tasks using standard benchmarks:

  • 2D Image Fitting: High-resolution RGB images (up to 8000x8000 pixels) to test representational capacity.
  • 3D Point Cloud to Mesh: Five meshes from the Stanford 3D Scanning Repository (e.g., Bunny, Dragon).
  • 3D Scene Reconstruction & Novel View Synthesis: The NeRF Synthetic and DTU datasets, which are the gold standards for evaluating how well a model can reconstruct a 3D scene from a few photos.

Key Insights and Ideas

  • The Power of Simplicity: Moving from a complex multiresolution hierarchy to a single adaptive grid simplifies the math for mesh extraction. It avoids the "cross-level intersections" that make previous tetrahedral methods (like TetraSDF) slow.
  • Locality is King: Because it is hash-free, every vertex in the tetrahedral grid has "local support." This is the "aha moment" for editing: you can grab a part of the mesh and move it, or change its color, and be 100% certain that you aren't breaking the rest of the scene.
  • Exact Correspondence: TetraCoder doesn't just approximate a mesh; it is a mesh. The implicit neural field and the explicit tetrahedral scaffold are two sides of the same coin.
Fig. 2. Tetrahedral ray tracing. For a sample 𝑥𝑗along the ray 𝑟(𝑡) = 𝑜+𝑡𝑑, an adjacency walk (purple) locates the containing cell. Shading highlights the visited cells, while green arrows indicate face normals. Dashed lines illustrate barycentric interpolation of vertex features within the target triangle (𝑎,𝑏,𝑐). Gray markers denote additional samples along the ray.

Key Results

  • Speed: TetraCoder achieves a 6.4x speedup in analytic mesh extraction compared to TetraSDF.
  • Accuracy: In 2D image fitting, it outperformed Instant-NGP in PSNR (image quality) while using fewer parameters.
  • Mesh Quality: It achieved the lowest SSDF (a measure of how well the extracted mesh matches the underlying field) compared to classical methods like Marching Cubes or Dual Contouring, even at extremely high resolutions.
  • Editing: The authors demonstrated that they could perform geometry deformation and appearance editing (like "painting" on a 3D object) without retraining the model, thanks to the localized nature of the encoding.
Fig. 3. R3 – Analytic extraction and grid-based extractors on the same frozen Lego field. SSDF measures mesh–field agreement.

Limitations and Future Work

While powerful, TetraCoder currently focuses on static scenes. The authors identify extending this to dynamic 4D scenes (video/motion) as a key next step. They also suggest that future versions could support multi-level analytic extraction, allowing for "Level of Detail" (LOD) rendering where distant objects use simpler meshes. Finally, they aim to further prove the "bidirectionality" of their link—showing that you can not only go from field to mesh but also seamlessly update the field by modifying the mesh in a standard 3D modeling tool.

Summary updated

Abstract

Implicit and explicit representations of geometry are central in computer graphics applications and come with their pros and cons. With surface sampling being faster on explicit representations and spatial queries being faster on implicit representations, finding efficient conversions between them is a highly desirable property. We present TetraCoder, a single resolution hash-free adaptive tetrahedral encoding that links an explicit tetrahedral spatial encoder with an underlying implicit signed distance field (SDF). The resulting hybrid implicit-explicit space is continuous piecewise affine (CPWA) by design, allowing to obtain an exact correspondence between the implicit SDF and the explicit analytic mesh. TetraCoder's design comes with additional perks, in particular: (a) it scales better than multi-resolution encoding methods due to its single resolution optimizable grid, (b) unlike hash-based methods, it is not affected by collisions, (c) it enables fast tetrahedral ray-tracing, geometry and appearance editing, directly through its adaptive tetrahedral scaffold. We demonstrate the versatility of TetraCoder by showing an improvement over state-of-the-art methods on multiple tasks: 2D/3D fitting from images and point clouds, 3D novel view synthesis, 3D reconstruction, and geometry/appearance editing.
BibTeX
@misc{pietra2026tetracoder,
  author = {Daniele Della Pietra and Antonella Rech and Nicola Conci and Nicola Garau},
  title = {{TetraCoder: Adaptive Tetrahedral Encoding for Neural Fields}},
  year = {2026},
  eprint = {2611.18267},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18267},
}
Figure