WorksComputer Science
TetraCoder: Adaptive Tetrahedral Encoding for Neural Fields
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.
Introduction
A neural field is a function represented by a neural network that maps coordinates to quantities such as color, density, or signed distance. Formally,
where denotes the input coordinates, the predicted field values, and the learned parameters.
Neural fields support learning continuous geometry and view-dependent appearance from observations [26, 38, 39]. However, the usage of these reconstructions for geometric processing, simulation, and editing also benefits from explicit surfaces. For an SDF, the surface is defined by its zero-level set. Sampling-based mesh retrieval approximates this surface, with accuracy depending on the sampling resolution [23]. Continuous piecewise-affine (CPWA) fields enable analytic retrieval of the learned surface [20, 29], establishing a direct correspondence between implicit and explicit geometry with no sampling errors. The spatial encoding therefore influences not only fitting accuracy but also how the learned surface can be extracted and edited.
Existing encodings come with different compromises. Multi-resolution hash grids capture detail efficiently [27], but hash collisions can couple distant regions, and trilinear interpolation does not preserve piecewise-affine structure. Tetrahedral encodings enable analytic mesh retrieval provided that the downstream decoders don’t introduce nonlinearities, yet overlapping grid levels require costly geometric intersections [29]. Fixed tetrahedral encodings [19] cannot adapt their spatial domains during training, while mesh-based encodings [1, 43] depend on reconstructed surfaces.
We introduce TetraCoder, a single-level, hash-free adaptive tetrahedral encoding (Figure 1) that establishes a direct correspondence between the learned implicit surface and its explicit representation, while enabling editing through the encoder grid itself. Barycentric feature interpolation and ReLU geometry decoding preserve the CPWA structure, while vertex insertion, pruning, and movement adjust the density and placement of grid vertices to capture fine geometric and appearance details. The single grid enables analytic mesh retrieval without cross-level intersections, supports tetrahedral ray traversal, and provides geometry and appearance editing through its vertices.
Figure 1. We present TetraCoder, an adaptive tetrahedral encoding for neural fields. Multiresolution hash encoding (left) interpolates features on fixed spatial grids at multiple resolutions. Its trilinear interpolation does not produce a piecewise-affine encoding, making analytic mesh extraction impractical. To address that tetrahedral multiresolution hash encoding (center) subdivides these grids into fixed tetrahedra and uses barycentric interpolation to obtain a piecewise-affine encoding. TetraCoder (right) instead uses a single-level, hash-free encoding defined on an adaptive tetrahedral volumetric mesh. Combined with a piecewise-affine decoder, this representation enables fast analytic mesh retrieval. Unlike other encodings, TetraCoder supports novel view synthesis, scene reconstruction, 3D appearance and geometry editing within a unified framework.
We evaluate TetraCoder on image fitting, mesh reconstruction from point clouds, joint novel-view synthesis and mesh reconstruction from RGB images, and scene editing. Our results demonstrate better fitting with fewer parameters for the first two tasks and the best average geometry and rendering scores among evaluated baselines on NeRF Synthetic and DTU. For mesh reconstruction from geometric data, TetraCoder reduces mean analytic extraction time by a factor of approximately 6.4 compared with TetraSDF [29].
Our contributions are:
We present a new general encoder for neural fields that matches or outperforms multiresolution hash encoding on a series of tasks, with less trainable parameters.
Our encoder stores radiance and surface features locally in a single-level foam of adaptive tetrahedra, addressing the issues of fixed spatial grid and hash collisions.
TetraCoder builds a direct correspondence between the learned implicit surface and its explicit surface representations, enabling analytic mesh retrieval.
Related Work
Spatial encodings for neural fields. Neural networks often learn low-frequency components before capturing fine details, a tendency known as spectral bias [31]. For neural fields, this motivates enriching the decoder’s input with Fourier [36] or local learned features to better capture spatial detail. Local capacity can be concentrated where it is needed [9] through sparse voxel refinement and pruning [22], hierarchical octrees [35], or block decompositions that adapt to signal complexity [25]. Multiresolution grids offer a complementary way to capture detail, with hashing reducing their storage cost [27]. Beyond storage, the interpolation rule determines
how features are interpolated in space: barycentric interpolation on permutohedral lattices [32] or tetrahedral hash grids [29] preserve a continuous piecewise-affine feature field. Features can also follow reconstructed geometry through point-supported fields [41] or interpolation over a tetrahedralization derived from a point cloud [19]. These choices present different constraints: hashing can couple features at unrelated locations [27], fixed tetrahedralizations cannot redistribute vertices during training [19], and multiresolution tetrahedral encodings introduce additional extraction costs by requiring intersections between tetrahedra across multiple grid levels [29]. TetraCoder addresses these constraints by storing features without hashing on a single tetrahedral grid that adapts through vertex movement, insertion, and pruning.
Neural reconstruction and novel-view synthesis. Implicit neural representations learn continuous fields, allowing novel views to be synthesized by volume rendering using density and view-dependent radiance [26]. Since density alone does not uniquely define a surface, geometry can instead be represented through occupancy [28] or a signed distance function (SDF) whose zero-level set defines the surface [38, 44]. Combining SDF-based reconstruction with multiresolution hash encoding improves training efficiency and captures finer geometric detail [21, 39].
Scene properties can also be stored directly on explicit primitives, such as voxels [6], 3D Gaussians [17] and surfels [30]. Using planar Gaussians [13] or aligning them with a surface [11] makes these primitives better suited to surface reconstruction. Other methods optimize a mesh in-the-loop together with 3D Gaussians [10] or by constructing a connected mesh from triangle splats and progressively making its faces opaque [12]. Foam-based methods partition space into Voronoi regions [7, 8, 37] or tetrahedra [24], simplifying ray traversal. TetraCoder combines the connected spatial structure of foam-based rendering with SDF-based volume rendering by decoding features interpolated over an adaptive tetrahedral grid. Neural decoders map the vertex features interpolated by TetraCoder to continuous SDF and radiance fields, allowing geometry and appearance to vary within each tetrahedron.
Surface representations and analytic extraction. Analytic surface mesh retrieval instead aims to retrieve the mesh that exactly matches the SDF. Marching Cubes [23] extracts an iso-surface from sampled grid values, while Dual Contouring [16] uses edge intersections and surface normals to guide vertex placement and capture sharp features. In tetrahedral grids, linear interpolation between scalar vertex values locates edge crossings that are connected into triangles by Marching Tetrahedra [5]. Making this extraction differentiable allows both vertex positions and SDF values to be optimized through losses on the resulting mesh [33]. Adaptive Delaunay grids extend this flexibility through error-driven resampling [3] and can also be paired with a separate hash-encoded SDF for image-based reconstruction [40]. Combining tetrahedral SDFs with cell rasterization further integrates differentiable extraction into rendering-based optimization [4]. For a linearly interpolated SDF, [5] recovers the represented surface exactly; a neural field, however, can contain details inside each tetrahedron that interpolation between its four vertex samples does not capture.
For ReLU MLPs, analytic extraction addresses this issue by identifying the regions in which the network is affine [2, 20, 34]. When a ReLU decoder operates on spatially interpolated features, the interpolation must also preserve piecewise-affine structure for analytic extraction to apply. Trilinear interpolation generally breaks this structure, motivating Eikonal-based planarity analysis and approximate intersection calculations [18], whereas barycentric interpolation preserves it when combined with a ReLU decoder [29]. However, using multiple tetrahedral grid resolutions requires resolving intersections between the grids and the decoder’s activation regions, making analytic extraction increasingly costly as the number of levels grows [29]. TetraCoder follows this analytic approach on a single adaptive tetrahedral grid, resolving the decoder’s ReLU activation regions to exactly recover the learned zero-level set and allowing it to scale better compared to current state-of-the-art methods [29, 34].
Editable neural fields. Editing a learned field often relies on an explicit geometric control structure, such as a mesh, rig, or cage. Mesh deformations can be transferred to a pretrained radiance field through a tetrahedral proxy [45], with enclosing cages extending this idea to coarse volumetric controls [42] and interactive manipulation [14]. An alternative is to make the control surface part of the representation itself by storing separate geometry and appearance codes at mesh vertices, supporting both deformation and local texture editing [43]. Local parameterizations and semantic guidance extend these controls and improve interaction efficiency [1]. Editing through a proxy relies on a separate deformation mapping, whereas storing codes on a surface mesh ties both the representation and its controls to that surface. TetraCoder uses the tetrahedral encoder itself, without needing an additional structure, as the editing domain: cage deformations move its vertices directly, while appearance edits modify its local features with the trained decoders fixed.
Method
This section is split into two parts. Section 3.1 describes the core structure, supported operations, and key properties of TetraCoder. All the presented details are meant to be agnostic to the task, and can apply to more tasks that go beyond the scope of this paper. Section 3.2 presents four scenarios that rely on our encoder to tackle increasingly difficult tasks. Task T1 (Toy 2D image fitting) shows that TetraCoder has enough capacity to fit an image into the weights of a small MLP, starting from 2D coordinates alone. Task T2 (3D point cloud to mesh) extends this idea to the 3D domain, where the MLP now stores a SDF in its weights. Being TetraCoder CPWA by design, we showcase its fast analytic mesh retrieval capabilities. Task T3 (3D scene reconstruction from images) demonstrates how TetraCoder jointly learns geometry and view-dependent appearance from a collection of RGB images. Its single-level adaptive tetrahedral grid supports both novel-view synthesis and explicit mesh reconstruction within the same representation, without additional spatial discretization, auxiliary 3D scaffolds, or mesh-in-the-loop learning. Task T4 (3D scene editing) proves this idea even further, showcasing the link provided by TetraCoder between the implicit SDF and explicit analytic mesh. Displacing the tetrahedral foam directly deforms geometry, while its hash-free design enables localized appearance edits without hash collisions causing changes in distant regions, thus allowing TetraCoder to become the editing domain itself.
Adaptive tetrahedral encoder
Grid and interpolation. TetraCoder partitions a 3D spatial domain using a single tetrahedral grid , where the single tetrahedron is defined as . The vertices of this grid are the corner points of its tetrahedra. Each grid vertex stores a position and a learned feature vector . To encode an input point , we first locate the tetrahedron containing it. Let denote its four grid vertices. For each vertex , we form by replacing its position with while keeping the other three vertices fixed. The barycentric weight of is the volume of this tetrahedron relative to the original volume:
where denotes tetrahedral volume. For a point inside , the four resulting volumes sum to the original volume, so the weights are nonnegative and sum to one. We then use these weights to interpolate the learned vertex features, obtaining the encoded feature vector:
CPWA structure and locality. The construction above gives TetraCoder two fundamental properties: a CPWA encoding and hash-free feature storage.
Continuity follows from interpolation over shared grid vertices. On a face between adjacent tetrahedra, the weights of the opposite vertices vanish, while the remaining weights coincide with the barycentric coordinates of that face. Both tetrahedra therefore interpolate the same features with the same weights, producing identical values at their interface.
Piecewise affinity follows from the volume-based definition of barycentric weights. Within a fixed tetrahedron , each volume in the numerator of Eq. (2) has a fixed base: the face opposite vertex . For inside , the height above this face varies affinely with . Since the denominator is constant, each barycentric weight is affine in . The encoding is a weighted sum of fixed vertex feature vectors, so it can be expressed by one affine function per tetrahedron:
Here, the matrix and offset are determined by the tetrahedron’s vertex positions and features and are constant within . Together with continuity across faces, this makes the encoding CPWA. Composing it with affine layers and ReLU activations preserves this structure, enabling analytic mesh retrieval from the learned field, as described in § 3.2.
The same vertex-based representation provides hash-free local support. Each grid vertex stores its own learned feature vector, which can be updated independently. For a fixed grid, changing one vertex feature produces
where is understood to be zero outside the tetrahedra containing . Changing the feature vector at vertex therefore affects the encoding only within tetrahedra incident to . With the decoder parameters fixed, modifying vertex features enables localized appearance editing without hash collisions causing changes in distant regions.
Grid adaptation and ray tracing. To represent coarse structures and fine details within a single encoding level, TetraCoder adapts local vertex capacity through densification, pruning, and vertex movement.
Densification back-propagates gradients to the interpolated features to identify regions requiring additional capacity. Let denote the primary fitting objective: the RGB reconstruction loss for T1 and T3, and the signed-distance loss for T2, defined in Eqs. (7) and (9), respectively. For each tetrahedron , let be a set containing the sampled points inside it whose interpolated features receive nonzero gradients. The refinement score is
where denotes the length of the longest edge of . The score is zero when is empty. This criterion favors spatially coarse regions with strong feature gradients. Insertion increases local vertex density and introduces independent feature vectors, initialized from the existing encoding as , where and denote the position and feature vector of the inserted vertex, respectively, and is the encoding before insertion.
Pruning identifies vertices whose features receive small training gradients over time and are similar to the mean feature vector of their neighbors. These vertices become candidates for removal. Removing selected vertices deletes their feature vectors and reduces local grid density.
Movement redistributes existing capacity by optimizing vertex positions as , where is an Adam update computed from gradients propagated through the encoder. Features remain attached to their vertices, so movement changes their spatial support without adding parameters.
Insertion and pruning trigger a full Delaunay rebuild over the active vertices. During movement, connectivity is refreshed periodically or when geometric checks detect inverted or excessively compressed tetrahedra. These refreshes update barycentric transforms and face adjacency while preserving surviving feature vectors.
The same connectivity supports tetrahedral ray tracing, our strategy for locating successive ray samples, illustrated in Figure 2.
Figure 2. Tetrahedral ray tracing. For a sample along the ray , 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 . Gray markers denote additional samples along the ray.
Tasks
T1 – Toy 2D image fitting. TetraCoder maps each normalized pixel coordinate to interpolated features . An RGB decoder predicts pixel colors as . We jointly optimize the encoder and decoder using the mean squared RGB loss:
where indexes the sampled pixels and is the reference RGB color of pixel .
T2 – 3D point cloud to mesh. TetraCoder supplies to a ReLU geometry decoder , defining the CPWA scalar field and the predicted signed distance . Given a batch of input points and reference signed distances , we jointly optimize the encoder and decoder with a main objective combining L1 fitting and Eikonal regularization:
where is a set containing the points used for Eikonal regularization, and the weight encourages unit SDF gradient norm.
Using the CPWA property described in § 3.1, TetraCoder supports analytic extraction directly over its single adaptive tetrahedral grid. We subdivide each tetrahedron into convex cells with fixed ReLU activation. The affine form in Eq. (4) can thus be extended to the coupled encoder and decoder with coefficients and . Since is zero exactly when its input is zero, the surface is equivalently defined by , giving within each cell
We triangulate the resulting nondegenerate planar polygons and assemble them into the surface mesh.
T3 – 3D scene reconstruction from images. TetraCoder forms the spatial representation of our image-supervised reconstruction pipeline, built on the volume-rendering framework of NeuS2 [39]. We feed to a ReLU geometry decoder following the formulation of T2, here using its CPWA scalar output directly as the SDF: . The geometry decoder produces the scalar SDF together with auxiliary features for appearance prediction ; we denote its full output vector by . A radiance decoder predicts local view-dependent radiance as
where is the viewing direction, and denotes the normal-related features evaluated at position . Let index sample points along a camera ray, ordered from near to far, and write . The opacity of the interval between and is
where is the logistic CDF and is a learned sharpness parameter. Volume rendering accumulates the local radiance along each ray, weighted by ray transmittance, to obtain the rendered pixel color .
We jointly optimize the encoder and decoders with the main objective
The learned SDF remains CPWA, enabling analytic mesh extraction as in T2.
T4 – 3D scene editing. Starting from a trained scene as in T3, TetraCoder supports geometry and appearance editing on the tetrahedral encoder grid, without retraining.
Geometry editing. TetraCoder edits geometry by moving its grid vertices while keeping the learned features, the tetrahedral connectivity, and the geometry decoder fixed. The learned field follows the grid deformation, producing a corresponding change in its zero-level surface and in the explicit mesh obtained through the analytic extraction described in T2.
Appearance editing. The geometry decoder produces both the SDF and . TetraCoder edits scene colors through its encoder grid. We perturb the features at selected grid vertices to match the feature vector that represents the desired color. The original features continue to determine the SDF and normals, preserving geometry while the edited features are used to render the edited color.
Experiments and Results
This section presents the averaged per-task results and a dedicated ablation studies subsection. We use Hash for multiresolution hash encoding, THash for tetrahedral multiresolution hash encoding and Mesh for methods that use a mesh-like scaffold for editing [1, 19].
R1 – Toy 2D image fitting
We evaluate image fitting on four high resolution RGB images at their original resolutions, up to 8000 8000 3 pixels. We optimize the weights of a single hidden unit (size 64) MLP to fit each image using Adam for 350 epochs and a learning rate 0.01. Each 2D sample is mapped to a 32-dimensional input vector using TetraCoder.
TetraCoder outperforms Hash [27] in terms of PSNR, while simultaneously reducing the number of required parameters
| Image | Method | PSNR (dB) ↑ | Parameters ↓ |
| Albert | Hash | 50.31 | 5.75M |
| TetraCoder | 59.81 | 3.61M | |
| Summer | Hash | 37.94 | 6.57M |
| TetraCoder | 38.18 | 4.82M | |
| Pluto | Hash | 44.03 | 23.96M |
| TetraCoder | 48.63 | 17.27M | |
| Girl | Hash | 35.96 | 86.20M |
| TetraCoder | 36.14 | 58.74M |
Table 1. R1 – Image fitting comparison with Hash Encoding (Instant-NGP). Our method consistently achieves higher PSNR with fewer parameters.
(Table 1). Qualitative results versus Hash are shown in Figure 4a.

Figure 4. (a) R1 – 2D Image fitting – Top-left: TetraCoder. Bottom-right: Hash. Compared to Hash, we obtain greater PSNR values with fewer parameters. (b) R2 – 3D Point cloud to Mesh – Details in red: TetraSDF artifacts due to the multi-level polyhedral extraction. Details in green: TetraCoder. (c) R3 – 3D Novel view synthesis – Details in green: TetraCoder can model view-dependent effects. Left: DTU Dataset. Right: Nerf Synthetic Dataset. (d) R4 – Editing – Left to right: TetraCoder fitted scene, appearance editing, geometric editing, and physics simulation involving another 3D object.
R2 – 3D point cloud to mesh
We evaluate analytic surface reconstruction on five meshes from the Stanford 3D Scanning Repository. For each normalized mesh, we generate 500,000 SDF supervision samples, with 95% perturbed surface samples and 5% uniformly sampled points in the unit cube . Surface samples are perturbed with isotropic Gaussian noise of standard deviation . Each query point is mapped by TetraCoder to an 8-dimensional feature vector and decoded by an MLP with three hidden layers of 12 neurons each. We train for 10 epochs using batches of 1,000 samples and Adam, with an initial decoder learning rate of 0.001 and cosine annealing. We report our results against TetraSDF Large [29] in Table 2, showing how TetraCoder can outperform tetrahedral hash encoding in terms of Chamfer Distance (CD), extraction time, number of parameters, SSDF and VSDF (as defined in [29]) while improving the output topology, with a greater number of Delaunay edges and less skinny triangles (Figure 4b).
| Metric | Method | Mean |
| CD ↓ | TetraSDF L (THash) | 1.718 |
| Ours (TetraCoder) | 1.611 | |
| Extract (s) ↓ | TetraSDF L (THash) | 2.859 |
| Ours (TetraCoder) | 0.446 | |
| Params. (M) ↓ | TetraSDF L (THash) | 16.778 |
| Ours (TetraCoder) | 5.221 | |
| Delaunay edges (%) ↑ | TetraSDF L (THash) | 82.37 |
| Ours (TetraCoder) | 83.50 | |
| Skinny triangles (%) ↓ | TetraSDF L (THash) | 19.55 |
| Ours (TetraCoder) | 9.14 | |
| SSDF ↓ | TetraSDF L (THash) | 0.076 |
| Ours (TetraCoder) | 0.027 | |
| VSDF ↓ | TetraSDF L (THash) | 0.078 |
| Ours (TetraCoder) | 0.027 |
Table 2. R2 – Surface reconstruction on the Stanford dataset. We compare reproduced TetraSDF (large) with our method using identical geometry supervision and 5,000 training updates. We report Chamfer Distance (CD, ), analytic extraction time in seconds, allocated parameters in millions, intrinsic Delaunay edge compliance (%), and skinny triangles (%).
R3 – 3D scene reconstruction from images
We evaluate both novel view synthesis from images and 3D analytic mesh reconstruction from the learned volume using PSNR, SSIM and CD. We encode 3D points to 28-dimensional features using TetraCoder and decode them using a decoder matching the structure of the NeuS2 [39] SDF and radiance decoders, training for iterations on NeRF Synthetic [26] and DTU [15]. The choice of the 28 features is not arbitrary: NeuS2 uses 14 levels of multi-resolution hash encoding with 2 features per level. Thus, we choose 28 features to match the same architectural capacity.
Table 3 shows TetraCoder consistently outperforming similar methods using Hash [27, 39], as well as recent methods specifically designed with additional editing scaffolds [1, 19]. Qualitative results are shown in Figure 4c.
Bridging the gap with task T2 (§ 4.2), we adapt TetraSDF [29] to fit the same image-based task. We first train this new architecture with a decoder matching the shape and size of TetraSDF L: a 4-levels encoder with a feature size of 2, combined with a (8, 128) MLP. We then train the same modified architecture but with a significantly larger, 14-levels NeuS2-style encoder and decoder, which we name TetraSDF-NeuS for convenience.
| Method | NeRF Synthetic | DTU | Mean | ||||||
| PSNR ↑ | SSIM ↑ | CD ↓ | PSNR ↑ | SSIM ↑ | CD ↓ | PSNR ↑ | SSIM ↑ | CD ↓ | |
| Tetra-NeRF [19] (Mesh) | 29.71 | 0.948 | 14.24 | 27.09 | 0.905 | 2.65 | 28.04 | 0.927 | 8.35 |
| NeuMesh++ [1] (Mesh) | 32.32 | 0.955 | – | 28.39 | 0.940 | – | 30.36 | 0.9475 | – |
| NeuS2 [39] (Hash) | 31.58 | 0.953 | 4.10 | 32.14 | 0.938 | 0.711 | 31.86 | 0.9455 | 2.4055 |
| Instant-NGP [27] (Hash) | 32.75 | 0.960 | 4.71 | 32.27 | 0.942 | 1.77 | 32.51 | 0.9510 | 3.2400 |
| Ours (TetraCoder) | 33.27 | 0.969 | 3.93 | 32.34 | 0.943 | 0.705 | 32.81 | 0.9560 | 2.3725 |
Table 3. R3 – Novel-view synthesis appearance and reconstruction accuracy on NeRF Synthetic and DTU. Mean reports the arithmetic average across the two datasets. CD is reported in millimeters.
Figure 3 shows a logarithmic plot comparing the analytic extraction time versus the SSDF for different methods. Classical methods (Marching Cubes, Marching Tetrahedra, Dual Contouring), show an exponential increase in extraction time when scaling the grid size. In the same plot, the modified TetraSDF L achieves similarly low SSDF to TetraCoder ( versus ), consistent with both methods using analytic extraction. However, it reaches only 23.53 dB PSNR and produces 93,065 boundary edges, 107,547 non-manifold edges and 26,374 duplicate triangles (all 0 for TetraCoder), showing that its limited capacity is insufficient to learn a high-quality SDF from images, yielding a severely defective mesh. At the same time, scaling TetraSDF-NeuS to NeuS2-like capacity makes construction of the polyhedral complex prohibitively expensive in memory and runtime, limiting its practical use at that scale. As shown in Figure 3, in less than 50 seconds TetraCoder achieves the lowest SSDF among all the candidates, almost two orders of magnitude lower than Dual Contouring trained on a grid for 2.5 hours. At resolution the other grid-based methods roughly match the resolution and storage size of our analytic mesh.
Figure 3. R3 – Analytic extraction and grid-based extractors on the same frozen Lego field. SSDF measures mesh–field agreement.
R4 – 3D scene editing
We evaluate 3D scene editing both in terms of geometry and appearance, and report the accuracy and the speed of our editing process. Table 4 quantifies unintended appearance changes by reporting the before–after MSE in unpainted regions. TetraCoder consistently preserves these regions better than Hash. These results demonstrate that local edits introduce limited changes outside the intended region.
| Editing % | 2D MSE↓ () | 3D MSE↓ | ||
| TetraCoder | Hash | TetraCoder | Hash | |
| 20 | 0.048 | 0.536 | 0.000 | 0.055 |
| 40 | 0.040 | 0.655 | 0.001 | 0.139 |
| 60 | 0.043 | 0.942 | 0.001 | 0.091 |
| 80 | 0.160 | 2.221 | 0.003 | 0.155 |
Table 4. R4 – Before–after editing MSE on the unpainted region for 2D and 3D scenes at different amounts of editing.
Figure 4 illustrates appearance editing, geometric deformation, and interaction with another object through physics simulation. Compared with NeuMesh++ [1], TetraCoder achieves approximately 2× the rendering speed and 19× and 18× lower geometry and appearance editing latencies, respectively.
Ablation studies
We ablate densification (D), pruning (P), and site movement (M) on NeRF Synthetic (Table 5). The complete configuration performs best across all metrics. Densification mainly supports appearance quality, while pruning and site movement primarily improve geometric accuracy. Their removal causes substantial geometric degradation with limited photometric changes, suggesting that similar rendered images can hide different surfaces and emphasizing the importance of site organization beyond appearance fitting.
| D | P | M | CD ↓ | PSNR ↑ | SSIM ↑ |
| ❌ | ❌ | ✓ | 7.185 | 31.5678 | 0.959210 |
| ❌ | ✓ | ❌ | 7.929 | 31.2010 | 0.955343 |
| ❌ | ✓ | ✓ | 9.036 | 31.5586 | 0.959147 |
| ✓ | ❌ | ✓ | 6.534 | 33.1690 | 0.968326 |
| ✓ | ✓ | ❌ | 7.094 | 33.2307 | 0.968409 |
| ✓ | ✓ | ✓ | 3.926 | 33.2647 | 0.968904 |
Table 5. Site-update ablation on NeRF Synthetic. D, P, and M denote densification, pruning, and site movement, respectively; ticks and crosses indicate enabled and disabled operations. We report mean CD (, ↓), PSNR (dB, ↑), and SSIM (↑).
Fig. 4. Qualitative results on T1: Toy 2D image fitting, T2: 3D point cloud to mesh, T3: 3D scene reconstruction from images, T4: 3D scene editing. TetraCoder can adapt to different decoders, tackling multiple tasks with a single unified representation.
Conclusion
We presented TetraCoder, a single-resolution, hash-free adaptive tetrahedral encoding that bridges implicit and explicit representations of geometry. By combining learnable vertex features with an adaptive spatial discretization, our approach efficiently allocates representational capacity while avoiding the limitations of multi-resolution grids and hash collisions. Its continuous piecewise-affine structure enables exact analytic mesh extraction, establishing a direct correspondence between the learned implicit surface and its explicit representation. The presented results establish TetraCoder as a versatile spatial encoding that unifies neural field learning, explicit surface reconstruction, and scene manipulation within a single adaptive representation. To address the current limitations, future work will extend TetraCoder to dynamic 4D scenes, enable analytic mesh extraction at multiple levels of detail, and prove the bidirectionality of the implicit-explicit link by recovering the learned SDF from the extracted mesh.
LLM Usage Statement. No LLMs or generative AI tools were used for research ideation, the formulation of hypotheses or scientific conclusions, or dataset creation and annotation. We used OpenAI ChatGPT and Codex to obtain reviewer-style feedback on the manuscript and to assist with drafting and editing text, refining related-work and citation material, reviewing and refining implementation code, and supporting experimental design, execution, and analysis. The authors take full responsibility for all submitted content, including the accuracy of the claims, citations, implementation, and reported results.
Acknowledgments
This research was supported by grants from NVIDIA and utilized two NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition GPUs.
References
References
- [1]Chong Bao, Yuan Li, Bangbang Yang, Yujun Shen, Hujun Bao, Zhaopeng Cui, Yinda Zhang, and Guofeng Zhang. 2025. NeuMesh++: Towards Versatile and Efficient Volumetric Editing with disentangled Neural Mesh-based Implicit Field. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025).arxiv.org/abs/2606.19316
- [2]Arturs Berzins. 2023. Polyhedral Complex Extraction from ReLU Networks using Edge Subdivision. In Proceedings of the 40th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 202). PMLR, 2234–2244.arxiv.org/abs/2306.07212
- [3]Alexandre Binninger, Ruben Wiersma, Philipp Herholz, and Olga Sorkine-Hornung. 2025. TetWeave: Isosurface Extraction using On-The-Fly Delaunay Tetrahedral Grids for Gradient-Based Mesh Optimization. ACM Transactions on Graphics 44, 4 (2025). doi:10.1145/3730851DOI
- [4]Jinkai Cui, Kaiwen Song, Chumeng Niu, and Juyong Zhang. 2026. Distance Field Rasterization for End-to-End Mesh Reconstruction. In ACM SIGGRAPH 2026 Conference Papers.
- [5]Akio Doi and Akio Koide. 1991. An Efficient Method of Triangulating Equi-Valued Surfaces by Using Tetrahedral Cells. IEICE Transactions on Information E74-D, 1 (1991), 214–224.
- [6]Sara Fridovich-Keil, Alex Yu, Matthew Tancik, QinHong Chen, Benjamin Recht, and Angjoo Kanazawa. 2022. Plexonels: Radiance Fields Without Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 5501–5510.
- [7]Shrisudhan Govindarajan, Daniel Rebain, Dor Verbin, Kwang Moo Yi, Anish Prabhu, and Andrea Tagliasacchi. 2026. Power Foam: Unifying Real-Time Differentiable Ray Tracing and Rasterization. arXiv preprint arXiv:2604.24994 (2026).arxiv.org/abs/2604.24994
- [8]Shrisudhan Govindarajan, Daniel Rebain, Kwang Moo Yi, and Andrea Tagliasacchi. 2025. Radiant Foam: Real-Time Differentiable Ray Tracing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 4135–4145.arxiv.org/abs/2502.01157
- [9]Shrisudhan Govindarajan, Zeno Sambugaro, Ahmedkhan Shabanov, Towaki Takikawa, Daniel Rebain, Weiwei Sun, Nicola Conci, Kwang Moo Yi, and Andrea Tagliasacchi. 2024. Lagrangian hashing for compressed neural field representations. In European Conference on Computer Vision. Springer, 183–199.arxiv.org/abs/2409.05334
- [10]Antoine Guédon, Diego Gomez, Nissim Maruani, Bingchen Gong, George Drettakis, and Maks Ovsjanikov. 2025. MILo: Mesh-In-the-Loop Gaussian Splatting for Detailed and Efficient Surface Reconstruction. ACM Transactions on Graphics 44, 6 (2025), 223:1–223:15. doi:10.1145/3763339DOI
- [11]Antoine Guédon and Vincent Lepetit. 2024. SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 5354–5363.arxiv.org/abs/2311.12775
- [12]Jan Held, Sanghyun Son, Renaud Vanegden, Daniel Rebain, Matheus Gadelha, Yi Zhou, Anthony Cioppa, Ming C. Lin, Marc Van Droogenbroeck, and Andrea Tagliasacchi. 2026. MeshSplatting: Differentiable Rendering with Opaque Meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 7320–7329.arxiv.org/abs/2512.06818
- [13]Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2024. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery. doi:10.1145/3641519.3657428DOI
- [14]Clément Jambon, Bernhard Kerbl, Georgios Kopanas, Stavros Diolatzis, Thomas Leimkühler, and George Drettakis. 2023. NeRFshop: Interactive Editing of Neural Radiance Fields. Proceedings of the ACM on Computer Graphics and Interactive Techniques 6, 1 (2023).
- [15]Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanaes. 2014. Large Scale Multi-view Stereopsis Evaluation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition. 406–413. doi:10.1109/CVPR.2014.59DOI
- [16]Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. 2002. Dual Contouring of Hermite Data. ACM Transactions on Graphics 21, 3 (2002), 339–346. doi:10.1145/566654.566586DOI
- [17]Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 2023. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics 42, 4 (2023).arxiv.org/abs/2308.04079
- [18]Jin-Hwa Kim. 2024. Polyhedral Complex Derivation from Piecewise Trilinear Networks. In Advances in Neural Information Processing Systems, Vol. 37.arxiv.org/abs/2402.10403
- [19]Jonas Kuhlánek and Torsten Sattler. 2023. Tetra-NeRF: Representing Neural Radiance Fields Using Tetrahedra. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 18458–18469.
- [20]Jiabo Lei and Kui Jia. 2020. Analytic Matching: An Analytic Meshing Solution from Deep Implicit Surface Networks. In Proceedings of the 37th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 119). PMLR, 5789–5798.arxiv.org/abs/2002.06597
- [21]Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. 2023. Neuralangelo: High-Fidelity Neural Surface Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 8456–8465.arxiv.org/abs/2306.03092
- [22]Lingjie Liu, Jiatao Guo, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. 2020. Neural Sparse Voxel Fields. In Advances in Neural Information Processing Systems, Vol. 33.
- [23]William E. Lorensen and Harvey E. Cline. 1987. Marching Cubes: A High Resolution 3D Surface Construction Algorithm. ACM SIGGRAPH Computer Graphics 21, 4 (1987), 163–169. doi:10.1145/37402.37422DOI
- [24]Alexander Mandel, Trevor Hedstrom, George Kopanas, Janne Kontkanen, Falko Kuester, and Jonathan T. Barron. 2026. Radiance Meshes for Volumetric Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 8267–8277.
- [25]Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. 2021. ACORN: Adaptive Coordinate Networks for Neural Scene Representation. ACM Transactions on Graphics 40, 4 (2021), 58:1–58:13. doi:10.1145/3450626.3459785DOI
- [26]Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. 2020. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision (ECCV).arxiv.org/abs/2003.08934
- [27]Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. 2022. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Transactions on Graphics 41, 4 (2022), 102:1–102:15. doi:10.1145/3528223.3530127DOI
- [28]Michael Oechsle, Songyou Peng, and Andreas Geiger. 2021. UNISURF: Unifying Neural Implicit Surfaces and Radiance Fields for Multi-View Reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 5589–5599.arxiv.org/abs/2104.10078
- [29]Seonghun Oh, Youngjun Uh, and Jin-Hwa Kim. 2026. TetraSDF: Analytic Isosurface Extraction with Multi-resolution Tetrahedral Grid. In Proceedings of the 19th European Conference on Computer Vision (ECCV).arxiv.org/abs/2511.16273
- [30]Hanspeter Pfister, Matthias Zwicker, Jeroen van Baar, and Markus Gross. 2000. Surfels: surface elements as rendering primitives. In Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH ’00). ACM Press/Addison-Wesley Publishing Co., USA, 335–342. doi:10.1145/344779.344936DOI
- [31]Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. 2019. On the Spectral Bias of Neural Networks. In Proceedings of the 36th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 97). PMLR, 5301–5310.
- [32]Radu Alexandru Rosu and Sven Behnke. 2023. PermutoSDF: Fast Multi-View Reconstruction with Implicit Surfaces Using Permutohedral Lattices. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).arxiv.org/abs/2211.12562
- [33]Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. 2021. Deep Marching Tetrahedra: A Hybrid Representation for High-Resolution 3D Shape Synthesis. In Advances in Neural Information Processing Systems, Vol. 34.
- [34]Christian Stippel, Felix Mujkanovic, Thomas Leimkühler, and Pedro Hermosilla. 2025. Marching Neurons: Accurate Surface Extraction for Neural Implicit Shapes. ACM Transactions on Graphics 44, 6 (2025), 222:1–222:12. doi:10.1145/3763328DOI
- [35]Towaki Takikawa, Joey Litallien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. 2021. Neural Geometric Level of Detail: Real-Time Rendering with Implicit 3D Shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).arxiv.org/abs/2101.10994
- [36]Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. 2020. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. In Advances in Neural Information Processing Systems, Vol. 33.arxiv.org/abs/2006.10739
- [37]Bernardo Taveira, Carl Lindström, Joakim Johander, and Fredrik Kahl. 2026. Differentiable Voronoi Ray Tracing Beyond Rasterization Speeds. arXiv:2608.17682 [cs.CV] https://arxiv.org/abs/2608.17682
- [38]Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. 2021. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction. In Advances in Neural Information Processing Systems.arxiv.org/abs/2106.10689
- [39]Yiming Wang, Qin Han, Marc Habermann, Kostas Daniilidis, Christian Theobalt, and Lingjie Liu. 2023. NeuS2: Fast Learning of Neural Implicit Surfaces for Multi-view Reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 3295–3306.arxiv.org/abs/2212.05231
- [40]Johannes Weidenfeller, Shaofei Wang, Philipp Fürstner, and Siyu Tang. 2026. ADELE – Adaptive Delaunay Grids for High-Fidelity Mesh-Native Reconstruction. In SIGGRAPH Asia 2026 Conference Papers. doi:10.1145/3829340.3842214DOI
- [41]Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. 2022. Point-NeRF: Point-Based Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 5438–5448. doi:10.1109/CVPR52688.2022.00536DOI
- [42]Tianhan Xu and Tatsuya Harada. 2022. Deforming Radiance Fields with Cages. In European Conference on Computer Vision (ECCV).arxiv.org/abs/2207.12298
- [43]Bangbang Yang, Chong Bao, Junyi Zeng, Huajun Bao, Yedia Zhang, Zhaopeng Cui, and Guofeng Zhang. 2022. NeuMesh: Learning Disentangled Neural Mesh-based Implicit Field for Geometry and Texture Editing. In European Conference on Computer Vision (ECCV).arxiv.org/abs/2207.11911
- [44]Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. 2021. Volume Rendering of Neural Implicit Surfaces. In Advances in Neural Information Processing Systems, Vol. 34. 4805–4815.arxiv.org/abs/2106.12052
- [45]Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. 2022. NeRF-Editing: Geometry Editing of Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 18353–18364.arxiv.org/abs/2205.04978


