acceptodds
Under review as a conference paper at ICLR 2027

Learning-based Delaunay Tetrahedral Mesh Generation from Signed Distance Field

Abstract

Converting signed distance fields (SDFs) into tetrahedral meshes requires balancing geometric accuracy, element quality, and computational cost. We present a learning-guided framework that combines adaptive sampling with classical geometric construction to accelerate this process. A neural network predicts density fields from the input SDF, learning where to allocate surface and volume samples from density targets derived from reference tetrahedral meshes. Adaptive blue-noise sampling, Delaunay tetrahedralization, and marching tetrahedra then construct a mesh whose boundary follows the input field. A post-processing stage then fixes geometric and topological errors and improves element quality. To validate our approach, we evaluate accuracy, element quality, topology, complexity, and runtime. On four held-out datasets at grid resolutions from to , our method matches the geometric accuracy of CGAL while running up to faster, and every output mesh is closed, manifold, and free of slivers. Applied to SDFs produced by an image-to-3D generative model, it robustly and efficiently meshes shapes on which classical algorithms fail.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
tetrahedral mesh

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wnVjOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a learning-guided pipeline for converting gridded SDFs into adaptive tetrahedral meshes. A 3D U-Net predicts surface and volume sampling densities, supervised by CGAL meshes. Seeds are drawn by GPU blue-noise sampling, Delaunay-tetrahedralized, and clipped with marching tetrahedra, and a classical post-processing stage then repairs dents, spikes, non-manifold elements, and slivers. The key idea is to replace per-shape iterative refinement with a feed-forward seed initialization. On 797 shapes from four datasets, the method matches CGAL’s accuracy while running about 2× faster, and produces closed, manifold, sliver-free meshes, at the cost of 1.4–1.6× more tetrahedra. On TripoSG-generated SDFs, it meshes every shape, whereas CGAL often drops parts or fails.

Critical strengths
  1. The design cleanly separates learning from geometry: the network only predicts where to place seeds, while connectivity, boundary fitting, and element quality are handled by classical algorithms. This sidesteps the difficulty of learning discrete tetrahedral connectivity directly, and it yields practically valid output: every mesh is closed, manifold, and sliver-free, with zero failures across 2,787 cases. That matters for downstream simulation, where invalid meshes are unusable.

  2. The method is practically relevant to generative 3D pipelines. It works directly on SDFs, needs no changes to the generation stage, and generalizes without fine-tuning to SDFs generated by TripoSG. On those, it meshes every shape, while CGAL frequently drops disconnected parts or times out. This shows a concrete use case where the approach is more robust than the standard tool.

Critical weaknesses
  1. The learned component’s contribution is not isolated. The main novelty is a feed-forward seed initialization from a standard 3D U-Net, and the rest of the pipeline is classical. There is no comparison with the same pipeline using a non-learned density (uniform, curvature-based, or local feature size, as in Guo et al. 2016). Since post-processing takes about 70–75% of runtime and provides all validity guarantees, the gains may come mostly from blue-noise sampling and post-processing rather than learning. The model itself is also under-analyzed: its loss terms, branches, and KDE targets are not ablated, and density accuracy is never reported.

  2. The evaluation setting limits the conclusions. The pipeline works on a rasterized SDF grid, so it is adaptive only in element size, and fidelity stays capped by the grid. Accuracy is measured against the marching-cubes surface of the same grid rather than the ground-truth meshes, which hides errors at sharp features. The closest baselines are missing: sampling plus Delaunay with a hand-designed sizing field, extraction plus remeshing plus tetrahedralization, and isosurface stuffing. Several comparisons are also unbalanced: GPU versus CPU, untuned CGAL parameters (CGAL is also the teacher), and TetGen and Gmsh forced to keep the dense marching-cubes surface.

Questions for authors
  1. How does the same pipeline perform with a non-learned density (uniform, curvature-based, or local feature size), or with Poisson-disk sampling plus Delaunay as in Guo et al. (2016)? This is needed to isolate what learning contributes.

  2. Please compare against an extract–remesh–tetrahedralize pipeline (e.g., dual contouring or FlexiCubes, adaptive remeshing, then TetGen or fTetWild) and against isosurface stuffing.

  3. Please report accuracy against the original ground-truth meshes, including sharp-feature or Hausdorff metrics, not only against the marching-cubes surface of the input grid.

  4. Why must the geometric stages (projection, bisection, surface recovery) query the rasterized grid rather than a continuous SDF, especially for neural fields such as TripoSG’s?

  5. How do runtimes compare when baselines are hardware-matched, and how sensitive are CGAL’s speed and quality to its parameters? Can CGAL’s failures be avoided by supplying initial seed points?

  6. Please give a per-stage runtime breakdown, time distributions (median and max), and per-dataset results, especially for the zero-shot sets and the generated SDFs.

  7. How accurate is the predicted density itself, and which loss terms and branches (global, local, KDE targets) actually matter?

  8. Whether density prediction is purely local and deterministic, and how well it generalizes?

Reviewer_tYf2OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a learning-guided framework that accelerates SDF-to-tetrahedral-mesh conversion by predicting surface and volume sampling density fields. It also introduces a post-processing pipeline to repair geometric and topological errors in the resulting tetrahedral meshes. The method is evaluated on a collection of shapes.

Critical strengths
  1. A learning-based framework is proposed to covert SDF into tetrahedral meshes.

  2. A specific post-processing is proposed to repair geometric issues of the original learning model prediction.

  3. Experiments were conducted on multiple shape datasets to validate the method.

Critical weaknesses
  1. The use of “Generation” in the title is a bit misleading. The method converts SDFs into tetrahedral meshes using a learning-based network rather than learning a generative model.

  2. The method is primarily compared with the classical CGAL method in terms of runtime and accuracy on shapes from existing datasets and image-generated shapes. The most recent baseline is TetGen (2015), which raises concerns about whether the comparisons reflect the current state of the art. How does the proposed method compare with more recent tetrahedral methods such as [a] and [b]?

  3. Even among the existing baselines, the proposed method does not achieve competitive performance compared with TetGen (2015) and Gmsh (2009).

  4. Can the method be extended to SDFs of room-scale scene meshes?

References:

[a] Alexandre Binninger et al. TetWeave: Isosurface Extraction using On-The-Fly Delaunay Tetrahedral Grids for Gradient-Based Mesh Optimization. SIGGRAPH 2025.

[b] Yiwen Ju et al. Adaptive Grid Generation for Discretizing Implicit Complexes. SIGGRAPH 2024.

Questions for authors

In Fig. 7, CGAL fails to reconstruct substantial portions of the meshes. Can this be resolved by adjusting its thresholds? How do the other methods perform on these examples?

Reviewer_qA3POpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper generates adaptive tetrahedral meshes from SDFs by combining learned surface and volume sampling densities with blue-noise sampling, Delaunay tetrahedralization, and geometric post-processing. On mesh-derived inputs, it achieves accuracy comparable to CGAL with approximately 1.8–2.4× faster construction, while producing 1.4–1.6× as many tetrahedra. Cross-dataset and TripoSG evaluations demonstrate promising generalization and robustness.

Critical strengths
  1. The complete pipeline achieves a useful trade-off between accuracy, construction time, and mesh quality. Table 1 shows substantial acceleration at comparable accuracy to CGAL, with a higher proportion of outputs meeting the reported quality criteria.
  2. The evaluation covers multiple datasets and generated SDFs. Successfully processing all 55 generated shapes demonstrates promising applicability to complex geometries.
Critical weaknesses
  1. The ablations remove blue-noise sampling or post-processing while retaining learned density prediction, leaving the benefit of learning insufficiently isolated. Controlled comparisons against reasonable non-learned density strategies are needed to establish its added value.
  2. CGAL failures and missing components may depend on initialization. An official segmented-image example demonstrates improved component coverage with component-aware initialization, although this does not establish the outcome for these SDFs. A stronger initialization baseline, with its cost included, is therefore needed to substantiate the broader robustness claim.
Questions for authors
  1. With the downstream pipeline unchanged, how does learned density compare with uniform or SDF-based geometry-aware densities in accuracy, S_0, tetrahedron count, and runtime? Please distinguish allocation benefits under matched point budgets from benefits of the complete strategy.
  2. With surface initialization covering disconnected components and its cost included, do the differences in failure rate, component recovery, and runtime persist?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.