arxiv: 2604.23537 · v1 · submitted 2026-04-26 · 💻 cs.GR

Recognition: unknown

Distance Field Rasterization for End-to-End Mesh Reconstruction

Jinkai Cui , Kaiwen Song , Chumeng Niu , Juyong Zhang

Authors on Pith no claims yet

Pith reviewed 2026-05-08 04:57 UTC · model grok-4.3

classification 💻 cs.GR

keywords signed distance fieldmesh reconstructiontetrahedral rasterizationdifferentiable renderingmulti-view geometrysurface reconstructionalpha compositing

0 comments

The pith

Optimizing a continuous signed distance field over a tetrahedral grid allows rasterization-based rendering and direct mesh extraction without post-processing.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Rasterization methods deliver fast image synthesis but leave surface extraction to separate heuristic steps. Signed distance field approaches define clean surfaces yet depend on slow ray marching. SDFRaster closes the gap by optimizing an SDF across a Delaunay tetrahedral grid, rasterizing the tetrahedra with alpha compositing for efficient rendering, and folding differentiable Marching Tetrahedra into the loop so meshes emerge directly from training. On standard multi-view benchmarks this produces more complete surfaces at lower memory cost than prior implicit or explicit techniques.

Core claim

SDFRaster represents a scene as a continuous signed distance field defined over a fixed Delaunay tetrahedralization. The field is optimized by rasterizing tetrahedra and alpha-compositing their signed-distance contributions; differentiable Marching Tetrahedra is integrated so that surface meshes are extracted as part of the same differentiable pipeline. This yields end-to-end mesh reconstruction from images that avoids post-processing extraction and attains higher completeness and accuracy with reduced storage on DTU and Tanks and Temples datasets.

What carries the argument

SDFRaster: a continuous signed distance field over a Delaunay tetrahedral grid, rendered via tetrahedra rasterization and alpha-compositing and coupled to differentiable Marching Tetrahedra for direct surface extraction.

Load-bearing premise

A fixed Delaunay tetrahedralization plus continuous SDF optimization and differentiable Marching Tetrahedra produces globally consistent surfaces without grid artifacts or any need for heuristic post-processing.

What would settle it

If SDFRaster meshes extracted on the DTU dataset show lower F-score or visibly more holes than competing methods such as NeuS or VolSDF, the higher-quality and completeness claims are falsified.

Figures

Figures reproduced from arXiv: 2604.23537 by Chumeng Niu, Jinkai Cui, Juyong Zhang, Kaiwen Song.

**Figure 1.** Figure 1: Our method introduces a rasterizable SDF representation for end-to-end mesh reconstruction from multi-view images. With this representation, our view at source ↗

**Figure 2.** Figure 2: Overview of SDFRaster. We learn a continuous SDF on a Delaunay tetrahedral grid, using a shared multi-resolution hash encoder to predict SDF values at vertices and appearance per tetrahedron. We render the images by rasterizing tetrahedra and alpha-compositing SDF-derived opacities. We apply differentiable Marching Tetrahedra on the tetrahedral grid with the learned SDF values to extract meshes in the opti… view at source ↗

**Figure 3.** Figure 3: Surface Reconstruction on the Tanks and Temples Dataset [Knapitsch et al. 2017] and Mip-NeRF 360 Dataset [Barron et al. 2022]. Qualitative comparison on four scenes (Barn, Truck, Bicycle, Kitchen) with MILo [Guédon et al. 2025], 2DGS [Huang et al. 2024], and our method. Our method converges to more accurate geometry while keeping the extracted meshes compact. Reference Image MC on uniform grid MT on our ad… view at source ↗

**Figure 4.** Figure 4: Comparison between Marching Tetrahedra (MT) [ view at source ↗

**Figure 5.** Figure 5: Comparison between 2DGS [Huang et al. 2024] and PGSR [C hen et al. 2025], which use depth fusion for mesh extraction, and our SDF-based mesh extraction. 2DGS and PGSR render multi-view depth maps and fuse them with TSDF [Curless and Levoy 1996], a pipeline that is sensitive to view-dependent depth noise and often struggles with thin structures. In contrast, we extract the zero level set from the learned SD… view at source ↗

**Figure 6.** Figure 6: Surface Reconstruction on the DTU Dataset [Aanæs et al. 2016]. Qualitative comparison on four scans (Scan 37, Scan 65, Scan 118, Scan 122) with MILo [Guédon et al. 2025], 2DGS [Huang et al. 2024], and our method. and recall at the standard distance threshold, and for DTU we follow the official evaluation script and report the average of accuracy and completeness as Chamfer distance. We cull extracted mesh… view at source ↗

read the original abstract

Rasterization based methods have recently enabled high-quality novel view synthesis at real-time rates, but their underlying volumetric primitives do not expose a direct, globally consistent surface representation, leaving sur face extraction to heuristic post-processing. In contrast, implicit signed dis tance field (SDF) methods provide well-defined surfaces but are typically optimized with computationally expensive ray marching. We propose SD FRaster, a rasterizable SDF representation that bridges this gap by combin ing the efficiency of rasterization with signed distance field for end-to-end mesh reconstruction. Starting from a Delaunay tetrahedralization, we op timize a continuous SDF over a tetrahedral grid and render it efficiently by rasterizing tetrahedra and alpha-compositing their contributions. We further integrate differentiable Marching Tetrahedra into the optimization loop, enablingend-to-endmeshreconstructionwithoutpost-processingmesh extraction. Experiments on DTU and Tanks and Temples demonstrate that SDFRaster achieves higher-quality and more complete surface reconstruc tions with lower storage cost than state-of-the-art approaches. Project page: https://ustc3dv.github.io/SDFRaster/

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SDFRaster rasterizes a continuous SDF over a fixed Delaunay tet grid and folds differentiable Marching Tetrahedra into the loop for end-to-end reconstruction, which is a clean technical bridge but leaves the method dependent on the starting grid's coverage.

read the letter

SDFRaster lets you optimize a continuous signed distance field by rasterizing tetrahedra with alpha compositing and extracting the mesh differentiably, avoiding separate post-processing steps that usually follow rasterization methods. The core move is to treat the SDF as values on a tetrahedral grid that can be rendered efficiently like volumetric primitives while still supporting a well-defined zero level set. They start with a Delaunay tetrahedralization, optimize the vertex values, composite the tet contributions during rasterization, and backprop through Marching Tetrahedra to update everything together. That combination is the actual new piece; it is not just another implicit or rasterization paper but a specific way to make the two talk to each other in one optimization loop. The reported results on DTU and Tanks and Temples show more complete surfaces and lower storage than the baselines they compare against, which is the practical payoff they claim. The experiments appear to support the headline that you get both speed and surface quality without the usual trade-off. The soft spot is the fixed initial Delaunay grid. Because there is no adaptive refinement described, any large gaps or topology changes the target surface needs must already be spanned by the starting tets; the continuous SDF can only adjust values on existing vertices and faces. If the preprocessing grid is not dense enough in the right places, the method cannot introduce new connections outside that structure, and the completeness gains would then rest on an unablated preprocessing choice rather than the optimization itself. This is a real but bounded limitation rather than a fatal flaw. The paper is for graphics people who build 3D reconstruction pipelines and want faster training than ray-marching implicits while still getting clean meshes. A reader who already works with differentiable rendering or tetrahedral methods will see exactly where to plug this in. It deserves a serious referee because the technical steps are concrete, the claims are testable on public data, and the central idea holds together even if the grid-dependence needs more scrutiny in review.

Referee Report

2 major / 2 minor

Summary. The paper proposes SDFRaster, a rasterizable signed distance field (SDF) representation for end-to-end mesh reconstruction. It begins with a fixed Delaunay tetrahedralization, optimizes a continuous SDF over the tetrahedral vertices, renders via tetrahedra rasterization and alpha compositing, and integrates differentiable Marching Tetrahedra to extract meshes directly during optimization without post-processing. Experiments on the DTU and Tanks and Temples datasets report higher-quality, more complete reconstructions with lower storage costs than state-of-the-art methods.

Significance. If the quantitative results hold under full verification, the approach offers a practical bridge between the speed of rasterization pipelines and the global consistency of implicit surfaces, potentially reducing heuristic post-processing in 3D reconstruction and enabling more efficient differentiable rendering for graphics applications. The explicit combination of established components (Delaunay tets, alpha compositing, diff. Marching Tetrahedra) without circular parameter fitting is a strength.

major comments (2)

[Abstract] Abstract and method description (paragraph beginning 'Starting from a Delaunay tetrahedralization'): the central claim of 'more complete' reconstructions depends on the initial fixed tetrahedral grid being dense enough to support target topology; without adaptive refinement or topology-changing operations described, the zero level set remains constrained to existing tet faces and cannot fill large voids or introduce new connectivity, which is load-bearing for the Tanks and Temples results.
[Experiments] Experiments section (DTU/Tanks and Temples claims): the superiority in quality and completeness is reported without ablations on initial tet density, grid resolution, or failure cases where the fixed Delaunay leaves gaps; this undermines verification that the end-to-end pipeline (not preprocessing) drives the gains.

minor comments (2)

[Abstract] Abstract contains multiple typographical errors: 'dis tance' (should be 'distance'), 'sur face' (should be 'surface'), 'SD FRaster' (should be 'SDFRaster'), and 'enablingend-to-endmeshreconstruction' (missing spaces).
[Abstract] The abstract states 'lower storage cost' but does not quantify the storage metric or compare it directly to baselines in the provided text; this should be clarified with explicit numbers or a table.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below with clarifications and commit to revisions that strengthen the presentation of our method's assumptions and experimental validation.

read point-by-point responses

Referee: [Abstract] Abstract and method description (paragraph beginning 'Starting from a Delaunay tetrahedralization'): the central claim of 'more complete' reconstructions depends on the initial fixed tetrahedral grid being dense enough to support target topology; without adaptive refinement or topology-changing operations described, the zero level set remains constrained to existing tet faces and cannot fill large voids or introduce new connectivity, which is load-bearing for the Tanks and Temples results.

Authors: We agree that the completeness of reconstructions in SDFRaster is dependent on the initial Delaunay tetrahedral grid providing sufficient coverage and density to support the target topology. The grid is constructed once as a fixed preprocessing step from the input views and scene bounds to encompass the reconstruction volume, allowing the optimized continuous SDF and differentiable Marching Tetrahedra to extract surfaces at the zero level set within those tets. We do not claim adaptive refinement or dynamic topology changes. To address the concern, we will revise the abstract and method section to explicitly describe the grid construction process, state the assumption of sufficient initial density, and add a limitations discussion on cases where large voids or unsupported connectivity may not be recovered. revision: yes
Referee: [Experiments] Experiments section (DTU/Tanks and Temples claims): the superiority in quality and completeness is reported without ablations on initial tet density, grid resolution, or failure cases where the fixed Delaunay leaves gaps; this undermines verification that the end-to-end pipeline (not preprocessing) drives the gains.

Authors: We acknowledge that the current experiments do not include ablations on tetrahedral grid density or resolution, nor explicit failure cases for gaps in the fixed Delaunay grid. This limits the ability to fully isolate the contribution of the end-to-end SDF optimization and differentiable extraction from the preprocessing. In the revised manuscript, we will add these ablations (varying tet density and resolution) along with failure case examples to demonstrate the role of the differentiable pipeline in achieving higher quality and completeness. We will also clarify that the tetrahedralization is a non-learned, fixed preprocessing step, while the core optimization loop is end-to-end. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation combines independent established components.

full rationale

The paper's core pipeline begins with a fixed Delaunay tetrahedralization, optimizes a continuous SDF on its vertices, rasterizes tetrahedra via alpha compositing, and closes the loop with differentiable Marching Tetrahedra for end-to-end extraction. None of these steps reduce by construction to fitted inputs, self-definitions, or load-bearing self-citations; each draws from prior independent literature on tetrahedral grids, SDFs, and differentiable isosurfacing. Empirical gains are reported on external benchmarks (DTU, Tanks and Temples) rather than being forced by the method's own parameterization. The approach therefore remains self-contained against external validation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The method rests on standard domain assumptions in graphics about the suitability of Delaunay tetrahedralizations for continuous field approximation and the differentiability of marching tetrahedra; no new free parameters or invented entities are explicitly quantified in the abstract.

axioms (2)

domain assumption A Delaunay tetrahedralization provides a suitable discretization for representing and optimizing a continuous signed distance field
Invoked when the method starts from tetrahedralization and optimizes the SDF over the grid.
domain assumption Alpha-compositing of per-tetrahedron contributions accurately approximates the global SDF during rasterization
Central to the efficient rendering step described.

invented entities (1)

SDFRaster representation no independent evidence
purpose: A rasterizable signed distance field over tetrahedral grids that supports differentiable mesh extraction
New proposed construct introduced to bridge rasterization and implicit surfaces; no independent external evidence provided in abstract.

pith-pipeline@v0.9.0 · 5493 in / 1407 out tokens · 48471 ms · 2026-05-08T04:57:44.512484+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages

[1]

ACM Trans

TetWeave: Isosurface Extraction using On-The-Fly Delaunay Tetrahedral Grids for Gradient-Based Mesh Optimization. ACM Trans. Graph. 44, 4 (8 2025). doi:10.1145/3730851 SIGGRAPH 2025 issue. Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hed- man. 2022. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In 2022 IEE...

work page doi:10.1145/3730851 2025
[2]

ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia) 41, 6 (2022)

Reconstructing Personalized Semantic Facial NeRF Models From Monocular Video. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia) 41, 6 (2022). doi:10.1145/3550454.3555501 Stephan J. Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. 2021. FastNeRF: High-Fidelity Neural Rendering at 200FPS. In 2021 IEEE/CVF International ...

work page doi:10.1145/3550454.3555501 2022
[3]

ACM Trans

Radiant Foam: Real-Time Differentiable Ray Tracing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) . 4135–4145. Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. 2020. Implicit Geometric Regularization for Learning Shapes. In Proceedings of the 37th Interna- tional Conference on Machine Learning (Proceedi...

work page doi:10.1145/3763339 2020
[4]

Vbench: Comprehensive benchmark suite for video generative models

Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, Seat- tle, W A, USA, 20654–20664. doi:10.1109/CVPR52733.2024.01952 Xiaoyang Lyu, Yang-Tian Sun, Yi-Hua Huang, Xiuzhe Wu, Ziyi Yang, Yilun Chen, Jiangmiao Pang, and Xiaojuan Qi. 2024. 3DGSR: Implicit Surface...

work page doi:10.1109/cvpr52733.2024.01952 2024
[5]

In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022

Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Render- ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR). 16187–16196. Cheng Sun, Min Sun, and Hwann-Tzong Chen. 2022. Direct Voxel Grid Optimization: Super-fast Convergence for Radiance Fields Reconstruction. In 2022 IEEE/CVF Con- ference...

work page doi:10.1109/cvpr52688.2022.00538 2022