acceptodds
Withdrawn from ICLR 2027

QuadLink: Autoregressive Quad-Dominant Mesh Generation via Point-Relation Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The generation of production-ready quad-dominant meshes is a corner-stone of modern 3D content creation. Generating anisotropic quad-dominant meshes from point clouds is challenging, as existing methods are typically limited to producing either pure triangular meshes or pure quadrilateral meshes with isotropic densities. In this paper, we present QuadLink, a unified framework consisting of three stages for quad-dominant mesh generation by linking points into structured faces. QuadLink formulates polygonal mesh generation as a hybrid centroid-conditioned vertex linking model: it first predicts a unified set of anchors (vertices and face centroids), then learns centroid-conditioned links that associate vertices with face centroids, and finally assembles polygonal faces with a quad-first strategy guided by robust geometric verification strategies. This link-based formulation enables efficient generation of sparse and anisotropic quad-dominant meshes with coherent edge flow and meanwhile supporting hybrid polygonal topology. To construct training data for this model, we further introduce a Tri-to-Quad Operator that converts artistic triangle meshes into quad-dominant training data via global merge selection. Extensive experiments show that QuadLink produces production-ready quad-dominant meshes from point clouds and achieves improved geometric fidelity and topological quality compared to prior baselines. Our method natively supports hybrid polygonal topology, generalizing to arbitrary n-gon meshes without architectural changes.

Primary area:
generative models
Keywords:
3D Asset Generation, Polygonal Mesh, Autoregressive Generative Model

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_4QawOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper generates quad-dominant meshes from point clouds, meaning meshes made mostly of quads with a few triangles, laid out the way artists do. Instead of writing the mesh face by face, it generates the vertices plus one centroid per face, then links each centroid to its vertices. A transformer generates these points, a learned feature space pulls each face's vertices close to its centroid, and faces are then built one centroid at a time, quads first and triangles next, with geometric validity checks. For training data, the authors turn a large set of artist triangle meshes into quad-dominant ones by pairing neighboring triangles. They report better results than triangle-based generators and classic quad remeshers, both in metrics and in a user study.

Critical strengths
  1. A simple idea that fills a real gap. Artists mostly use quads, but current generators make triangles, and the one quad method pads every face to the same length. This paper instead generates the vertices plus one centroid per face and links them, so triangles and quads are treated the same way. This matters to me because the idea is clean, does not depend on face type, and could carry over to other mesh tasks.

  2. The results are clearly closer to artist layouts. Linking vertices to centroids in a learned space, not by plain distance, handles long thin faces well. The outputs look more like artist work than triangle generators or classic quad remeshers, and the artists in the user study rank it first, though see my question on how that study was run. This matters because good layout is what current generators lack.

Critical weaknesses
  1. The "production-ready" claim is not yet fully supported. The evidence for it is mostly indirect: the paper does not report how clean the generated meshes themselves are, and it is not stated whether the editing examples are shown on meshes the method generated. My concerns here are fixable, but since this is the paper's main claim, I do not think it is established yet, and this is the main reason for my rating.

  2. The comparison is hard to judge from the paper alone. Several parts of the evaluation, including how the baselines were run and how the user study scores were aggregated, are unclear or not fully consistent. Since these comparisons are the main evidence, I cannot yet tell how large the gains really are.

Questions for authors
  1. Each centroid builds its face on its own. What happens when two centroids pick the same or overlapping vertex set, and how often do duplicate, overlapping or flipped faces appear in the generated meshes? Relatedly, if recon rate is the share of centroids that end up with a valid face, what happens at the remaining ones and do they leave holes?
  2. Were the meshes in the editing examples (UV unwrapping, beveling, edge loops) generated by QuadLink, or taken from the converted training data?
  3. What settings and face limits were used for each baseline and remesher, and how many faces do the test meshes have? Why is MeshAnythingV2, listed as a baseline, missing from the main table? And is the face-assembly ablation run from ground-truth points or from generated points?
  4. How were the user study scores averaged, which methods did participants actually see, and did they know which result came from which method? Were the triangle baselines shown as converted quad meshes or as their original triangles?

If these are answered, I will be glad to raise my score.

Reviewer_DGHrOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

QuadLink generates quad-dominant meshes from point clouds through three stages: autoregressive prediction of vertices and face centroids, learned centroid–vertex linking, and face assembly with geometric checks. The point-based representation supports mixed face types without fixed-length face padding. The paper also introduces a tri-to-quad operator for constructing training data while retaining the source vertex layout and UV seams. The model is trained on 400K curated assets and evaluated on 100 test meshes, with ablations and a study involving 26 professional 3D creators.

Critical strengths
  1. The centroid conditioned linking paradigm seems to be an efficient low-poly topology construction method that could have much potential.

  2. I find the data curation useful, especially for retaining the source vertex layout and UV seams. However, QuadGPT already describes an ILP-based tri-to-quad operator with the same per-triangle matching constraint (Appendix A.1, Eqs. 12–13). I would like a more clear explanation of what is new here and evidence that these differences improve the curated data or downstream generation.

  3. The paper is generally well-written and easy to follow.

Critical weaknesses
  1. Within the abstract, the script states that "... as existing methods are typically limited to producing either pure triangular meshes or pure quadrilateral meshes with isotropic density...". At the same time, I am aware QuadGPT does support mixed Quad-Triangle outputs. Although in line 186-189, the paper admits that QuadGPT has the strength, I still find putting such an important statement that is ambiguous and not proper.

  2. The work SpaceMesh, although cited, is not compared or discussed in detail. As many of the current diffusion-based low-poly generation methods (Tripo P1/P2, Meshy SmartTopology) adopt SpaceMesh as their topology decoder, it is crucial that we get to know the key differences between your centroid-conditioned formation framework and the SpaceMesh one. One thing that needs to be pointed out is that the formulation of the equation residing in lines 240–243 seems to be similar to the one in SpaceMesh. Moreover, what can be more critical is that in works like Tripo P1/P2 or Meshy SmartTopology, the decoding input itself lies in a high-dimensional space, instead of being mapped from raw Euclidean space back to a high-dimensional space, where the information bottleneck could be more severe.

Questions for authors
  1. This work adopts separate tokenization/vocabulary space for each axis, and in Appendix B.6, the author tries to give some evidence to argue the unusual design choice. However, it is not easy for me to understand what the Figure 15 and Figure 16 are trying to convey. Can the authors make it more clear?

  2. Can the authors give a better and more clear comparison on the data curation with the QuadGPT one?

  3. What is the main technical advantage of centroid-conditioned linking over SpaceMesh’s learned connectivity representation?

Reviewer_4kfVOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes QuadLink, a three-stage framework that first predicts vertices and centroids, then models their relationships using contrastive learning, and finally assembles faces by progressively checking candidate faces against validation criteria.

Critical strengths

The three-stage design is well motivated, and the paper includes extensive experiments.

Critical weaknesses

The paper does not cite Nexus or provide a comparison with tripop2.0. The Retrieval Space section also lacks a comparison with spacetime distance.

Questions for authors

Why is spacetime distance not used in the Retrieval Space section?

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

This paper generates quad-dominant meshes from point clouds, targeting the sparse, anisotropic layouts artists actually make. An hourglass transformer predicts vertices and face centroids autoregressively as coordinate tokens. A fine-tuned Michelangelo encoder then embeds the points so each face's vertices cluster around its centroid, and faces are assembled per centroid, quads first with triangles as fallback. Training data come from a Tri-to-Quad operator over 400k licensed internal assets. On 100 held-out models, five triangle generators passed through that same operator are compared with three quad remeshers, and the method reports the best CD, HD, IoU and EFR among those rows. It also tops a 26-creator study on all four criteria. An appendix fine-tunes the model to produce Goldberg polyhedra.

Critical strengths
  • Centroid anchoring turns mixed-polygon generation into a grouping problem. Each vertex and centroid appears once, so no per-face padding and no face-type token, and learned retrieval clearly beats Euclidean retrieval. MeshFlow already learns connectivity from contrastive embeddings, so the new part is centroid-conditioned grouping.
  • The headline numbers favor the method: HD 0.056 and IoU 0.995 against 0.086 and 0.978 for the next best, and effectively a tie on CD with QuadWild. Against the three remeshers, mean study scores are roughly 2.0-4.9 times QuadWild's.
Critical weaknesses
  • The comparison supports the combined pipelines, not the source of the topology gap. All five generative baselines are triangle models put through the authors' own operator, and the paper says such outputs are hard to convert cleanly, so QR, EFR and preferences partly measure the operator. The nearest native quad generators are missing too: QuadGPT is excluded for lack of code, and Mesh-Pro is never discussed.
  • Production readiness is claimed without evidence on global assembly failures. Face candidates are checked independently per centroid, no intersection or edge-incidence check is documented, and the set-union update may drop identical faces, so it does not show duplicates survive. Allowing open and non-manifold assets is a fair defence for some of this, but not for unintended self-intersections.
Questions for authors
  • Could you run the same Tri-to-Quad operator on the test-set reference triangulations and report QR and EFR? Pre- and post-conversion numbers for the baselines would help too. I would not read that as an upper bound. Is there a reproducible, input-compatible QuadGPT checkpoint for this test set, and if not, what blocks it? Please also discuss Mesh-Pro, and say whether Point2Quad's supplied-vertex recovery is fairly comparable to your anchoring.
  • Does the implementation canonicalize face elements before the set-union update, and is there global conflict handling beyond the local checks? Could you report rates of unintended intersecting or coincident faces and unused vertices, with reference-asset context? Please also define the recon rate in Table 2 and whether the timing is per shape or aggregated.
  • What input surface, face budget and failure policy did Instant-Meshes, QuadriFlow and QuadWild get? These decide how much the geometric comparison means.
  • Please give the EFR sensitivity parameter and the dihedral threshold for reference hard edges, and confirm the metric is a mean alignment score rather than an edge proportion.

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

Related papers

Discussion (0)

Sign in to comment.