acceptodds
Under review as a conference paper at ICLR 2027

GrowBrep: Growing Closed B-Rep Shells One Face at a Time under Grammar Constraints

Abstract

Boundary representation (B-Rep) is the native format of computer-aided design (CAD), and generating it means producing continuous geometry and discrete topology at the same time, with topology as a hard constraint. A B-rep is usable downstream only if it closes into a solid, yet existing generators can only check closure after generation. We present GrowBrep, an autoregressive B-rep generator that makes closure a property of its sequence grammar. GrowBrep serializes a solid as a shell that grows one face at a time along a frontier of open edges. Each edge is introduced exactly once and closed exactly once by a second half-edge, and the sequence cannot end while any edge remains open, so every sequence the model is allowed to finish describes a closed, consistently oriented shell. The grammar does not constrain geometry, which the model still has to learn. Geometry enters the sequence through per-primitive autoencoders with a finite scalar quantization bottleneck. A decoder-only transformer generates the sequence, conditioned on a complexity token and optionally on an input point cloud, and illegal tokens are masked out at every decoding step. On ABC-1M, 81.7% of GrowBrep's generation attempts yield valid watertight solids, versus 49.0% for the strongest baseline, under one strict validity checker applied to every attempt of every method without rejection sampling. On the distribution-level metrics (COV, MMD, and JSD) it is comparable to the strongest baselines. Conditioned on point clouds, it produces a valid solid for 80.0% of the inputs, compared with 51.6% for the closest prior method on its own published test set, at lower cost per cloud. Drawing five independent samples per cloud and keeping the first valid solid raises validity to 95.8%.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
CAD Generation, Boundary Representation (B-Rep), Autoregressive Models, Grammar-Constrained Decoding, Topological Constraints

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_qHyJOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

GrowBrep proposes a new method for history free B-Rep generation in which the sequence output by the transformer creates a topologically closed shell by design. An innovative “grammar mask” keeps track of open halfedges and ensures that every halfedge receives one and only one mating halfedge. An ablation show this gives substantial gains in solid validity.

The method is shown to produce more complex bodies with higher validity than any of the “trained from scratch” baseline methods.

Critical strengths

The core idea of the paper, to design the sequence to ensure a topologically closed shell, makes a lot of sense. The results clearly demonstrate that approach delivers higher B-Rep validity. Low validity is the “achilles heel” of direct B-Rep generation.

Grammar mask is interesting and the ablation study proves that it adds value.

The details around the B-Rep checking (in the appendix) are very good. The authors have a good understanding of solid validity. Taking multiple samples and using the checker or CD to pick the best is a valid strategy. Good to see almost 90% for K=5

Paper is well written and avoids too much AI generated text (a refreshing change this year).

Critical weaknesses

Appendix A2 is for me one of the most impactful bits of the paper. Validity as a function of face count is the key metric for understanding direct B-Rep generation. GrowBRep beats all the baselines on this metric for sure. I would move this into the main paper if you have any extra space post review.

There are no comparisons with frontier LLMs. Sadly, these are now getting so good that “trained from scratch” methods are become unappealing. Hard to know how to compare with them for point cloud to B-Rep tasks. I believe they do well at image to B-Rep now.

The paper correctly states that the sequence ensures a topologically closed shell. The geometry/topology can still be inconsistent. Some snapping to work around this is described appendix B. I would consider having something in the introduction to acknowledge this.

Questions for authors

Interesting that you added a direction flag. I would have thought this could be worked out with geometric tests. Are there places where it actually helps? Does the post processing ever need to ignore the value from the ML.

Reviewer_qF3POpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces GrowBrep, an autoregressive model that generates B-Rep solids face by face using a frontier-growth serialization. A grammar mask tracks open edges and restricts invalid tokens during decoding, guaranteeing combinatorial closure, connectivity, and consistent orientation for completed sequences. Geometry is represented using quantized face and edge tokens. Experiments on unconditional and point-cloud-conditioned generation show substantially higher validity than existing direct B-Rep generation methods while maintaining competitive geometric quality.

Critical strengths
  1. The frontier-growth grammar provides a simple and effective mechanism for guaranteeing combinatorial closure, consistent orientation, and valid edge references during decoding.

  2. The evaluation applies the same strict OCCT validity check to all methods, and GrowBrep achieves substantial improvements in both build and validity rates.

  3. The method supports both unconditional and point-cloud-conditioned generation while maintaining competitive geometric fidelity and distribution-level quality.

Critical weaknesses
  1. Although the paper acknowledges that program-based CAD generation offers stronger validity-by-construction guarantees, it provides no empirical comparison with this paradigm. A comparison on a compatible subset, or an analysis of representational coverage, would better justify the proposed trade-off.

  2. The paper lacks a controlled ablation isolating the contribution of the frontier serialization and grammar mask. Therefore, it is unclear how much of the validity improvement comes from the grammar rather than the model, training recipe, or post-processing pipeline.

  3. The representation excludes internal voids, disconnected shells, and non-manifold structures, but the paper does not quantify how much real-world or ABC data falls outside these assumptions.

Questions for authors
  1. Can the authors compare GrowBrep with a program-based CAD generation method on a compatible subset, or quantify how many ABC shapes cannot be represented by common CAD programs?

  2. Can the authors provide an ablation separating the effects of frontier serialization, grammar masking, and post-processing on validity?

  3. How much does each stage of the post-processing ladder contribute to the final validity rate and runtime?

  4. What fraction of the original dataset is excluded because of unsupported topology, primitive limits, or the frontier-label limit?

Reviewer_cs9eOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes GrowBrep, an autoregressive B-Rep generation method that grows a shell face by face. It uses grammar-constrained decoding to guarantee topology closure, while geometry is generated through quantized tokens and post-processing. Experiments demonstrate substantially higher solid validity than the evaluated baselines in both unconditional and point-cloud-conditioned generation, while maintaining competitive geometric quality.

Critical strengths
  1. The paper presents a meaningful idea of incorporating topological closure constraints into the generation process. The proposed grammar provides a combinatorial closure guarantee by construction, rather than relying entirely on post-processing.
Critical weaknesses
  1. The experimental comparison omits several relevant recent B-Rep generation methods, including BrepARG, BrepGPT, and HiFi-BRep for unconditional generation. This makes it difficult to assess the competitiveness of GrowBrep against recent approaches.
  2. The writing is sometimes difficult to follow, and some statements are imprecise. For example, the introduction claims that "assembly constraints can be computed directly from the faces and their adjacency," which is an overly strong statement. Figures 1 and 2 are also visually cluttered.
Questions for authors
  1. Could the authors provide comparisons with more recent B-Rep generation methods, such as BrepARG, BrepGPT, and HiFi-BRep?
  2. Could the authors improve the readability and clarity of the manuscript, revise inaccurate or overly broad statements, and improve Figures 1 and 2 to better illustrate the proposed method?
1 comment

Thank you for raising your score. The new experiments are in Appendix A.3 (Table 8) of the revised PDF, and they will stay in the final version. We agree the two figures can be clearer and will keep working on them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.