acceptodds
Under review as a conference paper at ICLR 2027

MeshRepair: Autoregressive Topology Repair for Generated Artisan Meshes

Abstract

Recent artisan mesh generation methods produce meshes whose topology comes close to the preferences of human artists. But gaps to artist quality remain because generated meshes commonly contain holes, stray faces, and duplicated faces. Regenerating a mesh from scratch is costly, and a new draw may change the preferred topology entirely or introduce errors elsewhere. In this work we study artisan topology correction. The task is to correct the topology of a generated mesh (the draft) in place, toward the layout an artist would make. We propose MeshRepair for this task. It is a face-level correction network based on a prefix-LM. It reads the mesh bidirectionally and emits an edit sequence autoregressively. The sequence consists of a deletion phase followed by an addition phase. We propose the novel formulation of the deletion head which picks faces to delete autoregressively and the addition head that selects vertices to form a face to add. The addition head also adds vertices when needed. Untouched faces pass through unchanged, so the corrected mesh keeps the vertices and the layout of the draft. The draft enters as a prefix and only the edits are generated. A correction therefore takes far fewer decoding steps than generating every face again with an autoregressive mesh generator. We benchmark against repair tools on scripted defects and on held-out drafts from two mesh generators. Across both scripted defects and generated drafts, improves the quad-equivalent face F1, removes non-manifold faces, and cleans up the topology.

Primary area:
generative models
Keywords:
mesh generation, autoregressive, repair

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_B54DOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces MeshRepair, a prefix-LM for repairing generated mesh topology through autoregressive face deletion and addition. It bidirectionally encodes the draft, uses pointer heads to select faces and vertices, and generates new vertices when needed while preserving unedited connectivity. Experiments on scripted defects and generated drafts show improved face F1 and fewer non-manifold edges, with gains primarily driven by deletion.

Critical strengths
  1. Practical relevance and appropriate problem formulation. Repairing generated meshes through sparse face edits addresses a meaningful problem while preserving satisfactory regions that full regeneration could alter. The explicit edit sequence also makes the modifications inspectable.
  2. Well-motivated model design. Bidirectional prefix attention provides full draft context, while autoregressive decoding captures dependencies between edits. The specialized pointer heads appropriately distinguish face deletion from vertex-based face addition, and the ablation results support their advantage over coordinate-based prediction.
Critical weaknesses
  1. Limited realism: Evaluation drafts use ground-truth vertices and generators seen in training. Vertex errors in practical outputs may lead to incorrect edits. Tests with perturbed vertices and unseen generators are needed.
  2. Weak face addition: Gains appear largely driven by deletion; addition recall is only 0.093 on MeshCarve and 0.024 on LATO2. Boundary edges also increase on LATO2, suggesting incomplete restoration of missing connectivity. These results support the method as an effective cleanup stage more strongly than as a complete topology-repair solution.
Questions for authors
  1. Robustness to vertex errors. How does repair performance change when draft vertices are displaced or missing? Can you report results under controlled perturbations and clarify how often the NEW mechanism recovers missing vertices?
  2. Generalization across generators. Since both evaluated generators are represented in training, have you tested MeshRepair on drafts from an unseen generator? If not, what aspects of the method do you expect to transfer?
  3. Contribution of face addition. Using the same trained model, how do the final face F1 and boundary-edge statistics compare between deletion-only output and the full deletion-plus-addition output? This would clarify the net benefit of the addition phase.
  4. Performance across topologies. Can you stratify results by mesh complexity and, where well-defined, genus and number of boundary components?
5 comments

We thank the reviewer for the detailed questions.

W1 / Q1 / Q2. Vertex errors and unseen generators. For perturbed vertices, we displace 5% of the vertices at random by 1 to 5 voxels. MeshRepair raises face F1 from .827 to .843 on MeshCarve and from .281 to .325 on LATO2, and reduces the non-manifold edges per draft from 183.7 to 15.2 and from 406.2 to 20.2 (Section 4.4 of the revised PDF). For missing vertices, we punch holes into the MeshCarve drafts: connected face patches covering 5% of the faces are removed, and vertices whose faces are all removed leave the draft. MeshRepair recovers 50% of the missing vertices within one voxel on the MeshCarve drafts and 57% on the scripted-defect drafts. Around the holes that lost vertices, MeshRepair closes 60% of the open edges on the MeshCarve drafts and 68% on the scripted-defect drafts (Section 4.4 of the revised PDF). At the same time, it avoids closing intended openings that can only be semantically determined: on the drafts of Table 1, its output covers 11.4% of the boundary edges of the reference (the openings the artist made) for MeshCarve and 0.6% for the scripted defects, fewer than the draft itself and than every hole-filling tool (trimesh, Blender, PyMeshLab). The details are in Table 4 of the revised PDF. For unseen generators, we evaluate MeshRepair qualitatively on meshes from Tripo and Meshy, i.e., commercial models, which generate their own vertex positions from an image. We include the visual samples in the revised PDF (Appendix D). Overall, MeshRepair repairs the topology of these meshes: it repairs overlapping faces and removes redundant faces, i.e., the repair ability is transferable.

W2 / Q3. Addition and boundary edges. The addition phase gives a net benefit: for the same model, the full deletion-plus-addition output has higher face F1 and fewer boundary edges than the deletion-only output on both generators (Table R1). Regarding the addition recall, the edit metric of Table 1 requires an exact triangle match: when an artist quad is split along the other diagonal, the surface is the same, but the addition is counted as wrong. Scored consistently with the face-set and edge columns (Table 5 in the revised PDF), the addition recall of MeshRepair is .299 on MeshCarve and .138 on LATO2. In addition, we believe that the addition recall itself is a limited measure of the additions, and we regard it as a reference only. In the hole-punching experiment above (W1 / Q1 / Q2), MeshRepair closes most of the open edges around the holes that lost vertices, and a closure that uses triangles other than the artist's can be equally valid, but under the exact match the addition recall does not count it.

Table R1. Same model, Table 1 drafts. Face F1 and boundary edges as a share of all edges.

DraftsMetricDeletion onlyDeletion + addition
MeshCarveFace F1.861.868
MeshCarveBoundary edges14%5%
LATO2Face F1.334.340
LATO2Boundary edges16%7%

Q4. Stratified results. We stratify the Table 1 drafts by the reference mesh. MeshRepair has the highest face F1 gain in every stratum on all three sets (Table R2).

Table R2. Face F1 gain over the draft, ours / best repair tool (MeshFix excluded, as in Table 1). Genus is computed on the edge-manifold references.

Reference meshMeshCarveLATO2Scripted
Faces <= 1,000+.011 / -.001+.079 / +.000+.082 / +.000
Faces 1,001-2,500+.027 / +.008+.026 / +.006+.081 / +.002
Faces > 2,500+.030 / +.008+.007 / +.006+.078 / +.004
Genus 0+.023 / +.007+.038 / +.005+.084 / +.001
Genus 1+.025 / +.007+.036 / +.006+.083 / +.003
Genus >= 2+.020 / +.008+.030 / +.002+.082 / +.002
Boundary loops 0+.024 / +.002+.068 / +.000+.084 / +.008
Boundary loops 1-5+.020 / +.007+.032 / +.007+.082 / +.003
Boundary loops > 5+.016 / +.005+.032 / +.007+.075 / -.003

Dear Reviewer,

Thank you again for your detailed review. As the discussion period closes on Nov. 18, we would greatly appreciate it if you could take a look at our response and the revised PDF. If our response addresses your concerns, we would be grateful if you could consider updating your rating. We are happy to clarify any remaining questions before the deadline.

I thank the authors for the additional experiments, particularly the deletion-only comparison. However, the presented examples appear to focus on relatively simple, local defects, whereas missing connections across large planar regions are, in my experience, a major challenge in AI-generated meshes. Without evidence on such cases, I remain uncertain about the method’s practical utility and maintain my Weak Rejection.

Reviewer_sx6oOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents MeshRepair, a model that repairs mesh topology through autoregressive face deletion and addition. It uses a bidirectional mesh prefix to understand the input, then selects faces to delete and vertices to form new faces. The model can also introduce new vertices while preserving untouched connectivity.

Critical strengths
  • The paper addresses a practical limitation of mesh generation: a mesh may have a desirable overall layout but still contain local topology errors. Repairing these errors while retaining the existing layout is a meaningful goal.
  • The editing formulation is well-motivated. Predicting explicit edits is a natural approach to local repair.
  • This paper is well-written and easy to follow.
Critical weaknesses
  • Predicting connectivity on the reference vertices is a useful controlled setting, but it does not establish performance on ordinary generated meshes with imperfect vertices or on unseen generators. These tests would better support the claim that MeshRepair can serve as a general repair stage.
  • Face recall changes very little on generated drafts. Under the reported edit metric, addition recall is only 0.093 on MeshCarve and 0.024 on LATO2.
Questions for authors

Please see the weaknesses.

2 comments

We thank the reviewer for the positive assessment.

W1. Generalization. For unseen generators, we evaluate MeshRepair qualitatively on meshes from Tripo and Meshy, i.e., commercial models, which generate their own vertex positions from an image. We include the visual samples in the revised PDF (Appendix D). Overall, MeshRepair repairs the topology of these meshes: it repairs overlapping faces and removes redundant faces, i.e., the repair ability is transferable. For imperfect vertices, we displace 5% of the vertices at random by 1 to 5 voxels. MeshRepair raises face F1 from .827 to .843 on MeshCarve and from .281 to .325 on LATO2, and reduces the non-manifold edges per draft from 183.7 to 15.2 and from 406.2 to 20.2 (Section 4.4 of the revised PDF).

W2. Face recall and addition recall. The face F1 gain is led by precision, since MeshRepair deletes faces the reference does not contain, while face recall is preserved (Table 1). Regarding the addition recall, the edit metric of Table 1 requires an exact triangle match: when an artist quad is split along the other diagonal, the surface is the same, but the addition is counted as wrong. Scored consistently with the face-set and edge columns (Table 5 in the revised PDF), the addition recall of MeshRepair is .299 on MeshCarve and .138 on LATO2. In addition, we believe that the addition recall itself is a limited measure of the additions. In an experiment where we punch holes into the drafts (faces and vertices are removed to create holes), MeshRepair closes 60% of the open edges around the holes that lost vertices on the MeshCarve drafts and 68% on the scripted-defect drafts (Section 4.4 of the revised PDF). At the same time, it avoids closing intended openings that can only be semantically determined (Table 4 in the revised PDF). Visually, the surface is closed at these edges, but under the exact match, a closure that uses triangles other than the artist's is not counted by the addition recall. We therefore also provide qualitative samples for reference (Appendices D and E of the revised PDF).

My concerns have been addressed. Therefore, I keep my score.

Reviewer_QYtEOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents MeshRepair, an autoregressive framework for repairing draft meshes, including those produced by mesh generation models. Rather than regenerating an entire mesh, the model predicts sequences of face-deletion and face-addition operations, reducing the number of output tokens required for repair. The model is trained with teacher forcing under a prefix-language-model formulation, with supervision applied to the edit sequence. Experiments on scripted and generator-derived drafts demonstrate advantages over existing mesh repair tools.

Critical strengths
  1. Predicting deletion and addition operations is an intuitive formulation that aligns with common mesh-editing workflows. Compared with regenerating the entire mesh, this representation can substantially reduce output sequence length and improve efficiency.

  2. The paper addresses a practical limitation of current artistic mesh generators, whose outputs can contain holes, non-manifold edges, and other connectivity defects.

  3. The evaluation includes generator-derived drafts in addition to scripted corruptions, providing a more relevant assessment than synthetic corruption experiments alone.

Critical weaknesses
  1. The preservation of intentional openings is not directly evaluated. An important motivation is that conventional tools cannot reliably distinguish accidental holes from intentional openings. Although the paper provides qualitative examples, it does not quantify how often MeshRepair incorrectly closes valid openings or connects structures that should remain separate. A targeted evaluation of unnecessary or harmful edits would strengthen this claim.

  2. More comprehensive qualitative comparisons are needed. It would be useful to show representative scripted drafts, MeshCarve-derived drafts, and LATO.2-derived drafts alongside the outputs of all repair baselines in Table 1. This would help readers understand which defects are corrected and which failure modes remain.

  3. Generalization beyond the generators used to construct the drafts remains unclear. The generator-derived training and evaluation drafts rely on MeshCarve and LATO.2 and use reference vertices. Consequently, these experiments primarily assess connectivity repair for known draft-generation pipelines rather than repair of unrestricted, independently generated meshes. Evaluation on an unseen generator, such as MeshFlow, without including its outputs in training would provide stronger evidence of generalization.

  4. The smaller advantage on generator-derived drafts deserves further analysis. Table 1 shows a more pronounced improvement over existing tools on scripted drafts than on generator-derived drafts. This may indicate a mismatch between scripted corruptions and generator-induced errors, although the results alone do not establish the cause. A model trained only on scripted drafts and evaluated on generator outputs, together with controlled experiments varying the training mixture, would help clarify the roles of corruption realism and data composition.

  5. The data-construction protocol needs additional detail. The paper should clearly identify the VAE used for reconstruction and explain how its outputs are converted into drafts with well-defined deletion and addition targets. For constructions using reference vertices, the authors should specify how those vertices are retained or imposed and how predicted connectivity is mapped to their indices. The total number of training examples and the proportions contributed by each construction procedure should also be reported.

  6. The scope of applicability beyond current generator-specific artifacts should be discussed. The motivation emphasizes holes, stray faces, and duplicated faces produced by existing mesh generators. As these generators improve, the frequency and distribution of such errors may change. The authors should clarify whether MeshRepair addresses a broader class of defects and how well it adapts to error patterns not represented during training.

Questions for authors

See weaknesses.

1 comment

We thank the reviewer for the constructive comments.

W1. Intended openings preservation. We count the boundary edges of the reference mesh (openings the artist made) that the output covers for every method. On all three sets, the output of MeshRepair covers fewer of them than the draft itself, while the hole-filling tools (trimesh, Blender, PyMeshLab) all cover more than the draft. The revised PDF adds this comparison (Table 4).

Share of reference boundary edges covered in the output (lower is better). MeshFix is excluded as in Table 1.

DraftsDrafttrimeshBlenderPyMeshLabOurs
Scripted4.1%16.6%46.0%86.6%0.6%
MeshCarve Generated16.4%24.3%62.4%93.3%11.4%
LATO2 Generated27.7%28.4%59.7%79.5%12.0%

W2. Qualitative comparison. The revised appendix adds side-by-side results of all repair baselines in Table 1 on scripted, MeshCarve and LATO2 drafts (Appendix E, Figures 5 and 6). Failure modes are also discussed.

W3. Generalization. We additionally evaluate MeshRepair qualitatively on meshes from Tripo and Meshy, i.e., commercial models, which generate their own vertex positions from an image. We include the visual samples in the revised PDF (Appendix D). Overall, MeshRepair repairs the topology of these meshes: it repairs overlapping faces and removes redundant faces, i.e., the repair ability is transferable.

W4. Gain on generated vs. scripted drafts. We agree that scripted damage may not fully represent error patterns of generative methods since the latter are more complex and noisier.

We evaluate an earlier version of the model that is solely trained on scripted damage and the face F1 results under the Table 1 protocol are shown as follows.

DraftsDraftMeshRepair w/o generator dataMeshRepair
Scripted (942).901.949 (+.048).981 (+.081)
MeshCarve (932).848.862 (+.014).868 (+.020)
LATO2 (938).294.302 (+.008).340 (+.046)

In sum, the scripted damage already offers valid supervision to the model though we recommend training the model with error patterns directly from generators for better performance.

W5. Data construction.

Our training data consists of two parts, scripted damage and generator outputs. The revised Appendix C gives these details.

  1. Scripted damage: we corrupt meshes in two ways. First, we punch holes into the drafts: connected patches of faces are removed, and the vertices whose faces are all removed leave the draft and are supervised as new vertices. In addition, the MeshCarve Edge VAE encodes each reference mesh and decodes it at noise levels in [0, 0.3], outputting meshes containing non-manifold edges, stray connections, etc., and the decoded mesh forms a reconstruction pair with its reference. Both holes and over-connections are therefore present in the training drafts by default.
  2. Generator outputs: MeshCarve and LATO2 generate connectivity on the vertices of training-split reference meshes, so that each generated draft is paired with its reference. The vertices are kept aligned so that valid supervision can be obtained. At training, we additionally punch holes on the surface.
  3. Supervision: in both parts, the damaged or generated mesh is the draft, and the artist mesh is the reference. The supervision signal is the edit sequence that turns the draft into the reference (Sections 3.1 and 3.2).
  4. Size and mixture: when training the base model, we draw from 524k artist meshes and inject the scripted damage at random. The fine-tuning set is obtained by applying the generators reported in the paper, i.e., MeshCarve and LATO2, to training-split reference meshes. During fine-tuning, the reconstruction, MeshCarve and LATO2 pairs are sampled at .50 / .25 / .25.

W6. Scope. MeshRepair is a condition-free topology repair model: it sees only the draft mesh, without any image or point-cloud condition. It therefore repairs topology errors of current diffusion and flow-matching models that can be semantically determined by the surrounding mesh, such as duplicated or overlapping faces, stray faces, and holes that can be recovered with semantic context (such as the topology around them). We also qualitatively validated MeshRepair on meshes from commercial closed-source models such as Meshy and Tripo, which have different frequencies and distribution of errors, and showed that it also generalizes to these samples. The results are included in Appendix D of the revised PDF.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.