acceptodds
Under review as a conference paper at ICLR 2027

CoMesh: Constrained Progressive Mesh Generation for Engineering Design

Abstract

Engineering shape design calls for models that explicitly generate meshes, which preserve broad geometric freedom and interface naturally with downstream physical simulations. Yet existing mesh generation approaches struggle to reconcile with spatial engineering constraints, achieve simulation-level quality, and often require costly autoregressive modeling over long refinement histories. We introduce *CoMesh*, a framework that starts with a prescribed seed mesh that hard-encodes engineering constraints and progressively introduces new geometric degrees of freedom through edge splits. Its key component is a novel serialization algorithm that constructs refinement trajectories for training in the same forward direction as generation, using a shared scheduler to select the edge refined at each step. We establish sufficient conditions for self-intersection free refinement, and translate them into a dihedral-based procedure for placing new vertices. Compared to baseline serializations, our approach guarantees that hard constraints can be embedded and reduces each refinement action to a 2D local displacement. This turns autoregressive generation into a local, state-conditioned problem, enabling an efficient, compact model independent of refinement history. Across three engineering families, mounting brackets, transonic wings, and ship hulls, CoMesh consistently generates diverse meshes with substantially improved constraint validity and element quality, while enabling markedly more efficient training and inference. These results demonstrate that using a deterministic refinement schedule with learned local geometry provides an efficient formulation for mesh-level generative design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Engineering design, Geometric constraints, Shape representation, Mesh generation, Progressive meshes

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

CoMesh is a framework for generating engineering surface meshes through progressive refinement of a prescribed, constraint-encoding seed mesh. Its main component, Forward Refinement Fitting (FRF), constructs training trajectories by refining the seed toward a target surface using the same deterministic edge-selection rule employed during generation. Each refinement action is represented as a two-dimensional displacement in an edge-local coordinate frame, or a stopping decision, enabling a compact model conditioned on local mesh geometry. The paper also derives sufficient conditions for self-intersection-free interior refinement and uses a simplified dihedral-based heuristic in practice. Experiments on mounting brackets, transonic wings, and ship hulls evaluate constraint preservation, mesh quality, diversity, and computational efficiency.

Critical strengths
  1. A well-motivated formulation that reduces the learning burden while preserving explicit constraints. By combining a prescribed seed mesh, deterministic edge selection, and fixed connectivity updates, CoMesh avoids learning operation locations and connectivity indices. Existing vertices are preserved throughout refinement, providing a direct mechanism for maintaining seed-vertex constraints. This decomposition is particularly suitable for engineering shape families with prescribed interfaces and fixed topology, and enables a compact local model.
  2. Substantial empirical improvements across multiple engineering shape families. The experiments show considerably fewer self-intersections and invalid outputs than several evaluated baselines, together with substantial reductions in generation time. The FRF/state versus PM/causal comparisons also provide useful evidence that the overall representation and modeling choices matter. These results suggest that the proposed formulation is a promising alternative to general-purpose mesh generators for the targeted constrained-design setting.
Critical weaknesses

1.The claim of stable mesh quality across resolutions needs qualification. Although the face-level self-intersection ratios remain low, Table 7 shows that the non-clear rate for wings increases from 3% at 768 faces to 16% at full resolution. Since this metric includes generation failures, degeneracy, and samples containing self-intersections, the increase raises a concern about sample-level reliability that is not adequately conveyed by the statement that mesh quality remains stable as resolution increases. Moreover, the bending-energy and normal-variation metrics in Appendix E.3 are resolution-dependent, so their decrease does not by itself establish improved geometric quality. A breakdown of non-clear outcomes and a clearer distinction between face-level quality and sample-level validity would help substantiate the scalability claim.

Questions for authors

Could you clarify how the claim of stable mesh quality across resolutions should be interpreted given that the wing non-clear rate in Table 7 increases from 3% at 768 faces to 16% at full resolution? In particular, could you break down these outcomes into generation failures, degeneracy, and self-intersections, indicating any overlap? This would help distinguish low face-level error rates from sample-level reliability and assess the practical implications of increasing resolution.

3 comments

We thank the reviewer for recognizing the value and methodological strengths of our approach. We also appreciate your careful reading of the resolution-scaling results, which has helped us clarify our claim and strengthen the evidence for practical use.

(1) Clarifying "stable mesh quality" and decomposing invalid outcomes. We agree that our original wording was insufficiently precise. Following your suggestion, we break down CoMesh's failure modes below and in the full-report tables in the manuscript. The results are updated to the mean of three training-seed experiments, so the values differ from the original report, but the trend remains the same.

Wing resolutionGeneration failure (%)With intersections (%)With degeneracy (%)Overlap (%)Intersecting faces (%)Invalid (%)
768 faces0.002.020.000.000.02842.02
1,920 faces0.002.150.000.000.02152.15
4,800 faces0.076.350.000.000.07826.41
Full resolution0.1012.430.000.000.142212.53

The invalid rate increases from 2.02% to 12.53% in the updated experiments and is primarily due to more intersections. The distinction between face-level and sample-level failures is also important: even a small fraction of intersecting faces can make an entire mesh unusable. Nevertheless, CoMesh retains a substantial advantage over the evaluated wing baselines, whose invalid rates exceed 50% even at 768 faces. Their reliance on global context at each generation step also makes scaling to higher resolutions difficult.

(2) Practical implications and an end-to-end route to physical analysis. To further demonstrate the usability of CoMesh-generated meshes, we have added downstream evaluations using the outputs directly in FEM/CFD analysis:

FamilyCompleted downstream stageSuccess rate
BracketAll three CalculiX shell load cases98.4%
WingVolume-mesh preparation for flow analysis86.3%
HullVolume-mesh preparation for flow analysis97.8%

Example simulation results are shown in Figures 24 (open the figure) and 25 (open the figure).

These experiments demonstrate how CoMesh's end-to-end generation of a design surface connects to physical analysis.

(3) Revising the role of bending energy and normal variation. We agree that their decrease with refinement alone cannot establish improved geometric quality. The revised main comparison therefore replaces the previous emphasis on smoothness with sliver-triangle ratio. We retain bending energy and the discussion of normal variation in the appendix as supplementary diagnostics. We have also added close-ups of bracket mounting interfaces (open the figure), hull deck alignment (open the figure), and wing triangle crossings (open the figure) in the revised Figure 6 (open the figure).

Thank you for the detailed response. The breakdown of invalid outcomes clearly shows that the increase at higher resolution comes mainly from self-intersections, and the added FEM/CFD downstream evaluations and the revised role of bending energy and normal variation address my concerns about how the quality claims should be interpreted. I would still encourage the authors to state explicitly in the main text that sample-level validity degrades with resolution (2.02% → 12.53% invalid for wings), and to list this as a limitation alongside the lower downstream success rate for wings (86.3%). Overall, I maintain my positive assessment and keep my score.

Thanks for your positive feedback. We will update the manuscript to explicitly state the numbers and clarify the limitations. As the first work targeting this constrained generation task, we believe these results already provide a promising starting point, with further improvements possible through parameter tuning.

Reviewer_akHLOpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents CoMesh, a constraint-preserving progressive mesh generation framework for engineering design. The method encodes spatial constraints in a prescribed seed mesh and incrementally generates a surface through edge refinement while preserving existing vertices. Experiments on mounting brackets, transonic wings, and ship hulls demonstrate advantages in constraint preservation, mesh validity, and computational efficiency.

Critical strengths
  1. The paper focuses on engineering-oriented shape generation rather than generic 3D object generation, providing a distinct and practically relevant problem setting.

  2. Preserving the prescribed structure of a seed mesh is a well-motivated objective, and the generation procedure is explicitly designed around this requirement.

Critical weaknesses
  1. The category coverage and evaluation scale are insufficient to establish broad generalization. The experiments cover only three relatively structured engineering surface families: brackets, wings, and hulls. It remains unclear whether CoMesh generalizes to objects with complex branching structures, holes, thin walls, or substantially different topologies. The authors should evaluate more challenging engineering shapes or explicitly delimit the supported seed structures and geometric configurations.

  2. The qualitative comparisons omit important references. Figure 5 and the additional comparisons in the supplementary material do not show the input seed meshes or corresponding reference meshes. This makes it difficult to visually verify constraint preservation and assess whether the generated connectivity is appropriate. A numerical table accompanying Figure 5 would also make quantitative differences easier to interpret.

  3. The advantage over existing mesh generators is not sufficiently established by the qualitative results. The outputs do not consistently show a clear improvement over MeshAnything V2 and TreeMeshGPT. A particularly relevant control would condition these autoregressive baselines on a prefilled mesh prefix representing the seed, where their representations permit this. Such a comparison would help determine whether CoMesh offers benefits beyond starting generation from an existing partial mesh.

  4. The method would benefit from an architectural overview. A diagram showing the inputs, conditioning information, model components, refinement process, and outputs would substantially improve clarity and make the intended task easier to understand.

  5. Control over engineering design parameters remains limited. The framework preserves seed constraints, but the paper does not demonstrate explicit control over properties such as curvature magnitude or the number and arrangement of curved regions. These capabilities would make the method more useful for engineering design. The authors should clarify which properties can be controlled through the seed and which remain unconstrained.

Questions for authors

See weaknesses.

5 comments

Thank you for recognizing the value of the constrained-design setting and seed preservation. Your suggestions have helped us clarify the method's scope and make the evidence easier to assess.

1. Broader applications (W1)

As one of the early efforts toward engineering mesh generation, our work covers three families to be the starting point, that already represent different application scenarios: overall shape design (wings and hulls) and local connecting-structure design (brackets). We agree that more examples are helpful to demonstrate our model's generalizability. We have therefore added two groups of cases to the appendix:

(1) Connecting multiple panels. We test FRF serialization on three- and four-arm brackets, with seed meshes constructed simply by connecting the corners of prescribed mounting panels, as shown in Figure 15 (open the figure). FRF successfully fits these samples, providing a basis for naturally extending CoMesh's training and application workflow to multi-branch connection tasks.

(2) Working with CAD tools. For more complex shapes, CoMesh can work with CAD tools operated by a human or a CAD agent to represent and design local surfaces within a larger model. The seed serves as an interface, where upstream CAD tools specify the overall structure, mounting panels, and edges connecting the surfaces, leaving local geometric details to model refinement. We provide two examples in the new Appendix C.4.2. Figure 16 (open the figure) shows local recess design: CAD defines a folded support with mounting ears and edge flanges, leaving a window for a recess that strengthens the structure. The rectangular interface defines a planar, two-triangle seed from which CoMesh represents the recessed surface through refinement. Figure 17 (open the figure) shows the design of a connecting part within an assembly: supports and rails define mounting panels (red) and upper rims (gold), while four additional routing points (purple) specify the main surface level of the part. Together, these form a seed for CoMesh to refine.

2. Better demonstration of our headline results (W2)

We thank the reviewer for these suggestions. We have made the following changes:

(1) A quantitative comparison table replaces the previous bar plot. It reports seed displacement, self-intersection ratio, invalid rate, number of components, mesh quality, and inference time. The reason for TreeMeshGPT's missing wing results is now stated explicitly, where its global quantization cannot distinguish nearby vertices at the trailing edge.

(2) Seed displacement is now reported relative to a reference length. (mounting-panel width for brackets, root chord for wings, and seed-defined deck beam for hulls.) Our CoMesh and the refinement baseline, VertexRegen, achieve exactly zero displacement, while other baselines show relatively large displacement, which will cost extra non-trivial postprocess to regain interface validity or for downstream simulation.

FamilyModelSeed (%) ↓SI (%) ↓Invalid (%) ↓Components → 1Slivers (%) ↓Time (s) ↓
BracketCoMesh0 ± 00.02 ± 0.0131.5 ± 0.731.00 ± 02.32 ± 0.26.9
VertexRegen S059.00100.01.5221.0829.3
MeshAnything V211.1140.3555.31.032.0059.0
TreeMeshGPT4.2591.7625.01.121.15699.3
MeshFlow (SG)11.32681.74100.0800.002.651.3
MeshFlow (CVPR)26.7010.9494.61.015.585.7
LATO.210.7490.5269.91.3310.376.0
WingCoMesh0 ± 00.03 ± 0.00622.0 ± 0.311.00 ± 00.84 ± 0.0444.1
VertexRegen S087.72100.01.0023.8138.9
MeshAnything V28.6570.5556.21.031.3357.9
TreeMeshGPTNot applicable—————
MeshFlow (SG)1.30674.96100.0768.001.201.2
MeshFlow (CVPR)1.1063.3899.51.003.755.1
LATO.21.1990.7685.01.1310.586.0
HullCoMesh0 ± 00.01 ± 0.00490.7 ± 0.391.00 ± 02.40 ± 0.1114.9
VertexRegen S0.83573.04100.01.1018.9441.7
MeshAnything V218.8610.2675.01.000.8167.4
TreeMeshGPT1.2910.5943.81.031.30105.9
MeshFlow (SG)8.23387.50100.01152.002.101.5
MeshFlow (CVPR)13.5611.1698.41.005.126.5
LATO.25.5652.7099.61.3912.766.1

CoMesh reports mean ± standard deviation across three training seeds. SI is the intersecting-face percentage; invalid is Fail + Non-clr. over all requests; slivers have minimum angle <10.

(3) Additional zoom-in plots in Figure 6 (open the figure) visualize seed displacement by comparing generated panels with the prescribed interfaces. In particular, the hull views show that baseline models do not consistently produce a deck boundary lying in a single plane, as required for downstream simulation.

(4) Further visualizations of all baseline outputs are provided in the appendix, with the seed meshes explicitly shown: brackets (open the figure), wings (open the figure), and hulls (open the figure).

3. Advantages over MeshAnything V2 / TreeMeshGPT (W3)

We address three aspects below and have updated the corresponding manuscript sections to convey these points more clearly.

(1) Prefilled seed prefixes for MeshAnything V2 and TreeMeshGPT. We would like to clarify that the seed already enters both baselines through a prefilled prefix. Their native encoders receive points sampled from the prescribed seed. Yet this prefix encodes the seed as a soft constraint since the prefix is a conditioning embedding derived from the seed mesh.

The reason we do not hard-encode the seed mesh, as CoMesh does, is that these baselines emit final-mesh triangles without operations for replacing existing faces. Our seed is a coarse mesh of the intended design, so generation requires refinement, rather than completion or expansion. Directly forcing the seed into the output prefix would retain faces that should be refined. Hard-encoding the seed does work for our refinement-based autoregressive baseline, VertexRegen, and this is exactly what we implement.

(2) Improvements in mesh validity and quality. Strict seed preservation is only one of CoMesh's advantages. CoMesh also improves mesh validity and quality, as shown below. On 32 bracket generation tasks with 32 samples each, CoMesh achieves an invalid rate of only 1.5%, with low self-intersection and clean connectivity. The two baselines have much higher invalid rates, despite larger training budgets and longer inference times. We also demonstrate downstream use by directly passing CoMesh outputs into FEM/CFD analysis: 98.4% of brackets complete CalculiX shell analysis, while 86.3% of wings and 97.8% of hulls complete volumetric mesh generation with Gmsh. Example simulation results are shown in Figures 24 (open the figure) and 25 (open the figure).

MetricCoMeshMeshAnything V2TreeMeshGPT
Seed-to-surface error (% panel width) ↓0.00011.1144.259
Intersecting faces (%) ↓0.020.351.76
Invalid generations (%) ↓1.555.325.0
Face-connected components → 11.001.031.12
Bending energy (smoothness) ↓1.61802.2610.94
Generation time (s) ↓6.959.0699.3

We attribute CoMesh's substantial improvement to its local prediction strategy, as discussed in Section 6.1. With forward refinement fitting (FRF), the model only needs to predict a local refinement decision on a selected edge. Replacing global prediction with local decisions greatly reduces the required model size and the information predicted at each step, while naturally providing invariance to scale, translation, and rotation.

(3) End-to-end design and mesh generation. CoMesh targets a different task from the two baselines who are primarily designed to reconstruct a mesh from an already specified target, whereas CoMesh jointly generates the design surface and its discretization end to end. To evaluate the baselines on their intended task, we also deliberately supply privileged target information in the GT-leak controls:

ModelInputFacesNME (%)SI (%)Slivers (%)ComponentsFail (%)Non-clr. (%)Seed (%)
CoMeshSeed7030.000.0192.321.000.031.50
VertexRegen 125MSeed prefix678.60.009.8511.371.510.0099.900
VertexRegen 125MGT cloud742.90.007.2011.201.030.00100.0010.034
TreeMeshGPTSeed cloud844.40.041.761.151.120.0025.004.259
TreeMeshGPTGT cloud786.30.050.571.351.0660.069.184.455
LATO.2Seed only953.527.421.868.031.000.00100.0010.851
LATO.2GT scaffold913.526.580.889.281.000.00100.0010.552

The results show that, even with the ground-truth point cloud as input, the baselines still perform worse than CoMesh in mesh validity and quality.

Reviewer_YZVROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces CoMesh, a framework for progressively generating triangle meshes from prescribed seed meshes that encode geometric interface constraints. Its main contribution is Forward Refinement Fitting (FRF), which constructs training trajectories in the same forward direction as generation and uses a shared deterministic edge scheduler, allowing a compact local-state model to predict only a two-dimensional vertex displacement or a stopping decision. The authors provide sufficient conditions for embedding-preserving interior refinement and evaluate the method on mounting brackets, transonic wings, and ship hulls through baseline comparisons, ablations, resolution-scaling experiments, and sampling analyses. The reported results demonstrate exact preservation of prescribed seed vertices, low self-intersection rates, and improved training and inference efficiency relative to the evaluated autoregressive baselines.

Critical strengths
  1. Forward Refinement Fitting (FRF) offers a meaningful methodological contribution by constructing training trajectories directly from prescribed seed meshes and sharing a deterministic refinement scheduler between training and generation. This reduces the burden of predicting refinement sites and connectivity while preserving prescribed seed vertices and frozen edges, providing a reasonable formulation for shape generation subject to interface constraints.

  2. The proposed representation enables a compact model to predict new vertices from the current local mesh state without explicitly processing the full generation history. The ablations across serialization strategies and model types, together with experiments on three geometry families, demonstrate the potential of this design to improve generation validity and computational efficiency.

Critical weaknesses
  1. The paper’s positioning as an engineering shape design method is not yet sufficiently supported by application-level evidence, and its role within a practical engineering workflow remains unclear. The experiments primarily evaluate geometric validity and surface smoothness, without establishing whether these improvements translate into better physical performance, simulation usability, or design efficiency. Although SuperWing and Ship-D provide relevant physical performance information, the paper does not establish a connection between the generated geometries and their engineering utility.

  2. If the main contribution is generating geometrically valid surface meshes under spatial constraints, its advantages over existing generation and post-processing pipelines require further substantiation. Although the comparisons include MeshFlow, they do not sufficiently clarify the specific benefits over diffusion-based 3D generation followed by mesh extraction and remeshing. These benefits should be evaluated in terms of interface preservation, connectivity, triangle quality, and post-processing costs under comparable input conditions and face counts. Establishing these advantages and identifying their sources are important for assessing the need for the proposed generation mechanism and the significance of its contribution.

  3. In a practical engineering workflow, who provides the seed mesh, how is it constructed, and what additional processing is required before the generated output can enter downstream design or simulation? Which existing operation does CoMesh replace or improve, and what evidence currently demonstrates the resulting benefit?

  4. In Algorithm 1, p0p_0, CC, and the argument rr in the SPLIT operation are undefined; there is a stray “E” at the end of the first paragraph of Appendix D.2; and the expression σqGT∈−1,+1\sigma_q^{GT}\in-1,+1 is missing set braces.

Questions for authors

Please refer to Critical Weaknesses.

2 comments

Thank you for recognizing our two key contributions, and we appreciate your comments and suggestions.

1. The role in an engineering workflow (W1 & W3)

In this work, we focus on learning geometric features and position CoMesh as an end-to-end shape representation under prescribed constraints, working with outer-loop physics solvers that provide functional feedback (open the figure). The general design workflow with CoMesh is:

Designer-specified geometric constraints (e.g., mounting panels) → a seed mesh constructed from a case-specific template → a solver-ready mesh generated by CoMesh refinement → a downstream physics solver.

This paradigm can be applied at multiple stages of an engineering workflow, with either human designers or design agents. The revised Figure 2 (open the figure) illustrates these workflows. Among the cases we demonstrate:

(1) Wing and hull design targets aero/hydro-dynamic shape optimization. A conventional workflow selects a geometric parameterization, determines its parameters, reconstructs the geometry, and then discretizes it for downstream solvers to obtain physical feedback. CoMesh offers a new route: an end-to-end refinement representation that combines these stages. For example, a designer specifies a wing planform, including its span and kink layout. A template constructs the seed, as shown in Figure 2 (open the figure), and CoMesh generates diverse sectional geometry through refinement. This provides a route to overcome the design-space restrictions imposed by a chosen parameterization and simplify gradient propagation across stages when optimizing parameters using physical feedback.

(2) For brackets, the upstream design specifies the positions and orientations of the mounting panels, and the goal is to design the connecting structure. To our knowledge, this task still relies largely on manual trial and error. CoMesh provides an end-to-end alternative: a simple template bridges the prescribed panel meshes to construct a seed, and CoMesh generates the connecting geometry and its mesh while preserving the prescribed interfaces exactly. We further demonstrate FRF on three- and four-arm connectors using similar templates in Figure 15 (open the figure) (Appendix C.4.1), showing that it accommodates branching structures.

(3) Beyond the three families, CoMesh can work with CAD tools operated by a human or a CAD agent to represent and design local surfaces within a complex model. The seed serves as an interface where upstream CAD tools specify the overall structure leaving local geometric details to model-based refinement. We provide two examples in the new Appendix C.4.2. Figure 16 (open the figure) shows local recess design: CAD defines a folded support with mounting ears and edge flanges, leaving a window for a recess that strengthens the structure. The rectangular interface defines a planar, two-triangle seed from which CoMesh represents the recessed surface through refinement. Figure 17 (open the figure) shows the design of a connecting part within an assembly: supports and rails define mounting panels (red) and upper rims (gold), while four additional routing points (purple) specify the main surface level of the part. Together, these form a 16-vertex, 14-triangle seed for CoMesh to refine.

2. Engineering relevance and application-level evidence (W1)

We have added direct downstream evaluations of CoMesh outputs for each family. 98.4% of brackets complete all three CalculiX shell load cases (open the figure). For wings and hulls, 86.3% and 97.8%, respectively, pass volume-mesh preparation for flow simulation. We use Gmsh to construct hybrid prism–tetrahedral meshes with 10 extruded layers. (open the figure) further presents two examples of OpenFOAM flow simulations on these meshes.

3. Advantages over generation followed by extraction and postprocessing (W2)

(1) We have expanded the flow-based baselines in our main comparison to include CVPR MeshFlow and LATO.2.

(2) We have tested native postprocessing and mesh reconstruction for two autoregressive baselines (TreeMeshGPT and MeshAnything V2) and two diffusion baselines (MeshFlow and LATO.2). We compare these results with CoMesh's end-to-end outputs below. The results show that postprocessing improves mesh quality and reduces non-manifold edges and disconnected components. However, postprocessing struggles to correct fundamental defects in the baseline outputs, such as large holes in the two diffusion models, spurious branches in TreeMeshGPT, and displacement from the prescribed interfaces in almost all models. Even after reconstruction, the results remain inferior to CoMesh's end-to-end outputs and fail to meet downstream simulation requirements.

ModelProcessingFacesNME (%)Slivers (%)ComponentsFail (%)Non-clr. (%)Seed (%)PP (s)
CoMeshUnprocessed7030.002.321.000.031.50.000—
MeshFlow (SIGGRAPH)Raw8000.002.65800.000.00100.011.326—
Native PP16000.002.64799.970.00100.011.2890.047
+ cleaning8000.002.64799.970.00100.011.2890.290
+ BPA reconstruction8000.004.105.500.004.717.8750.556
+ BPA reconstruction + cleaning8260.004.015.610.00100.017.8770.856
LATO.2Raw7664.2210.011.500.0068.810.609—
Native PP96127.379.951.380.00100.010.4760.015
+ cleaning5880.005.4411.690.00100.010.5000.234
+ BPA reconstruction7660.003.618.310.006.214.9470.530
+ BPA reconstruction + cleaning7850.003.538.750.00100.014.9820.805
TreeMeshGPTRaw8440.041.151.120.0025.04.259—
Native PP8440.041.151.120.0025.04.2590.005
+ cleaning8430.001.161.190.0025.04.2590.181
+ BPA reconstruction8440.005.093.880.006.211.0090.612
+ BPA reconstruction + cleaning8680.004.964.000.00100.011.0090.923
MeshAnything V2Raw7430.092.001.0330.1825.111.114—
Native PP (replay)7430.062.001.0330.1824.311.1140.112
+ cleaning7420.002.001.0930.1830.311.1160.272
+ BPA reconstruction7420.005.551.9830.181.216.1010.685
+ BPA reconstruction + cleaning7640.005.412.0330.1869.816.1010.957

NME: non-manifold edges. Non-clr.: intersecting or degenerate returned meshes. PP is cumulative postprocessing time. Native: each model's native postprocessor. Cleaning: removes duplicate/degenerate faces, splits disconnected vertex fans, repairs winding, and fills near-planar holes. BPA reconstruction: samples 4,096 area-weighted surface points and normals and reconstructs the mesh using the Ball-Pivoting Algorithm.

(3) Further tests with GT-point-cloud conditioning. We also test an alternative that can be viewed as a form of postprocessing: a GT-leak test in which baseline models are trained to generate meshes from point clouds sampled from reference GT meshes. This separates CoMesh's two roles—shape design and mesh generation.

The results show that, even when only mesh generation is required, the baselines still perform worse in mesh validity and quality. This further supports CoMesh's advantages in mesh generation through its compact representation and local features.

ModelInputFacesNME (%)SI (%)Slivers (%)ComponentsFail (%)Non-clr. (%)Seed (%)
CoMeshSeed7030.000.0192.321.000.031.50
VertexRegen 125MSeed prefix678.60.009.8511.371.510.0099.900
VertexRegen 125MGT cloud742.90.007.2011.201.030.00100.0010.034
TreeMeshGPTSeed cloud844.40.041.761.151.120.0025.004.259
TreeMeshGPTGT cloud786.30.050.571.351.0660.069.184.455
LATO.2Seed only953.527.421.868.031.000.00100.0010.851
LATO.2GT scaffold913.526.580.889.281.000.00100.0010.552

4. Undefined symbols and typo errors (W4)

Thank you for pointing out these errors. The intended notation is p0=me=(vl+vr)/2p_0=m_e=(v_l+v_r)/2; CC should be γe∗\gamma_e^*. The split uses the fixed midpoint ratio r=1/2r=1/2. The bending sign should read σqGT∈{−1,+1}\sigma_q^{\mathrm{GT}}\in\{-1,+1\}, and the stray “E” should be deleted.

Comments on the paper

3 comments

We thank the reviewers for recognizing the relevance of our constrained engineering design setting, the methodological contribution of FRF and its compact local-state formulation, and the improvements in constraint preservation, mesh validity, and computational efficiency. We also appreciate their careful assessment and constructive feedback. Before detailed responses, we want to state that our work explores a new direction in the area, and we view this as an early step toward directly connecting generative models with engineering design and physical analysis.

To help illustrate our responses, we provide anonymous links to figures from the manuscript, which you can click to view directly.

The main revisions are:

  1. Clearer task and scope. We revised the overview figure and added illustrations of the workflows and model architecture. We also added FRF demonstrations for branching connectors and CAD-defined local surfaces, and clarified the limitations of the current design controls.

  2. More informative metrics and visualizations. A quantitative table now replaces the headline bar plot, with normalized seed displacement and sliver-triangle ratio. We also added close-up comparisons and galleries of baseline outputs to make interface errors and local defects easier to inspect.

  3. Direct downstream evaluation. We provide structural analysis results and selected OpenFOAM flow simulations, demonstrating the CoMesh generated meshes can be directly used in downstream physical analysis.

  4. Broader comparisons and additional controls. We added CVPR MeshFlow and LATO.2, clarified how the baselines receive seed information, and evaluated postprocessing, mesh reconstruction, and privileged-input (GT-leak) controls. CoMesh retains a substantial advantage in mesh validity even when the evaluated baselines receive target geometry.

  5. More precise reliability claims. We report CoMesh results across three training seeds and decompose invalid outcomes by failure type. We clarify that sample-level invalidity increases with resolution, primarily because of intersections, and distinguish this from the fraction of intersecting faces.

Dear reviewers, we sincerely appreciate your assessment and constructive questions. We have updated the manuscript and provided detailed responses to address your concerns. We would be grateful for your feedback and are happy to provide further clarification if any concerns remain unresolved. Thank you.

Dear Reviewers,

The authors have submitted their rebuttal to address your concerns. As the discussion deadline approaches, please review their response and share any remaining concerns or questions soon so the authors have enough time to respond.

Thank you, -AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.