acceptodds
Under review as a conference paper at ICLR 2027

TOPOGEN: TOPOLOGY-GUIDED MULTIMODAL GENERATION FOR EXECUTABLE CADQUERY SYNTHESIS

11 views

Abstract

Multimodal models have shown promise in programmatic CAD generation; however, generating executable and structurally faithful programs from visual observations remains challenging. A model may reproduce the overall shape while failing to capture the structural relations and modeling operations required for engineering validity. Unlike many generative tasks, CAD provides an executable geometric representation that a geometric kernel can deterministically verify, making structural errors explicitly measurable. We present TopoGen, a topology-guided multimodal framework for CadQuery synthesis from point clouds and multi-view images. Our central insight is that CAD topology provides a shared structural space bridging heterogeneous visual observations and executable programs. TopoGen encodes topology as a constraint graph and derives geometry-grounded structural constraints from BReps for supervision during training. For execution-grounded reinforcement learning, we introduce Topology Relative Policy Optimization (ToRPO), which uses essential-operation matching to prevent reward exploitation and provide structural feedback during program generation. This design preserves both the geometric and topological relations of generated CAD models. We further construct TopoCAD-1M, a topology-augmented corpus that enriches large-scale mechanical engineering data with structural constraints and alleviates the long-tailed distribution of CAD operations.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Boundary Representation, CAD Program Synthesis, Constraint Learning, Reverse Engineering

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PBVrOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents TopoGen, a framework for generating executable CadQuery programs from point clouds or multi-view images. It uses BRep-derived constraint graphs for structural supervision and introduces ToRPO, which combines geometric rewards with essential-operation and feature checks. The authors also construct TopoCAD-1M, a CAD dataset annotated with constraint graphs. Experiments across multiple datasets show improvements in geometric reconstruction and program validity in most evaluated settings.

Critical strengths
  1. Significant problem. The paper addresses structural consistency in visual CAD reconstruction and uses BRep-derived constraint graphs to provide explicit supervision for program generation.
  2. Valuable dataset. TopoCAD-1M combines approximately one million CAD models with constraint annotations and covers a range of geometric complexities.
  3. Clear writing. The manuscript is well organized and clearly presents the research motivation and overall framework.
Critical weaknesses
  1. Limited novelty. The method follows an established SFT–RL framework with a GRPO-style objective, while its constraint-graph supervision offers limited novelty over prior structural supervision.
  2. Insufficient baseline comparisons. Only three baselines are evaluated per modality, with limited coverage of recent CadQuery-based methods. The paper also leaves unclear whether baseline results were reproduced or cited and why they differ from the original reports.
  3. Limited evaluation of complex models. The qualitative results provide limited evidence of performance on highly complex parts, and results across different complexity levels are not reported.
  4. Insufficient ablations. The ablations lack a baseline without topology supervision under the same RL setup and do not evaluate the essential-operation and feature rewards separately. The individual contributions of these components remain unclear.
  5. Insufficient dataset documentation and long-tail analysis. Operation frequencies and distribution comparisons are missing, leaving the long-tail mitigation claim insufficiently supported. CadQuery program acquisition and validation, along with training sample counts and sampling procedures for each modality, also remain unclear.
  6. Incomplete evaluation and reporting. The evaluation lacks direct measures of constraint and essential-operation recovery, limiting support for structural consistency. Mean and median CD are not reported separately, despite the statement in the text.
Questions for authors
  1. Component ablations. Please add a matched SFT + ToRPO baseline without topology supervision, compare geometric-only and full rewards, and evaluate the essential-operation and feature rewards separately.
  2. Dataset construction and long-tail distribution. Please provide operation-frequency distributions for the source data and TopoCAD-1M, together with performance on rare operations. Please also describe how the CadQuery programs were obtained and validated.
  3. Structural recovery. Please report quantitative results on constraint-relation and essential-operation recovery to establish whether improved geometric metrics also reflect better structural recovery.
  4. Input modalities and training setup. Please clarify whether both modalities can be input jointly and how they are used during SFT and RL, including sample counts and sampling proportions for each modality.
  5. Baseline results. Please specify the source, checkpoint, and evaluation settings for each baseline, and explain the differences from the original results.
Reviewer_3viNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TopoGen generates CadQuery programs from point clouds or multi-view images. It introduces BRep-derived constraint graphs and trains a multimodal model to generate the graph followed by CAD code. Topology Relative Policy Optimization (ToRPO) uses geometric and operation-based rewards. The paper also introduces TopoCAD-1M, a dataset with structural annotations and sampling organized by BRep face count. Evaluation covers conditioned reconstruction from images and point clouds on four benchmarks, with generally promising results, alongside a progressive ablation of the training stages.

Critical strengths

1. A useful structural-supervision formulation for executable CAD generation.

The paper addresses a meaningful distinction between final geometry and design-relevant structure. Extracting constraint graphs from BReps as intermediate supervision provides a concrete way to exploit existing CAD structure across both input modalities, without requiring ground-truth constraints at inference.

2. Promising results across both modalities.

Table 2 reports improvements over Cadrille on geometric and invalidity metrics, although not uniformly on every benchmark and metric. Table 3 also shows incremental improvements from topology-guided SFT and subsequent ToRPO. These results make the approach worth investigating.

Critical weaknesses

See the details below.

Questions for authors

1. The claimed graph-relevant reward is not in the ToRPO objective.

The introduction states that ToRPO augments geometric rewards with graph consistency, and Figure 2 explicitly includes "Topology: Graph matching." However, Eq. (6) defines only

R=0.70RIoU+0.10Ress+0.10Rfeat+0.05RCD+0.05RHD.R = 0.70R_{\mathrm{IoU}} + 0.10R_{\mathrm{ess}} + 0.10R_{\mathrm{feat}} + 0.05R_{\mathrm{CD}} + 0.05R_{\mathrm{HD}}.

The accompanying definitions describe geometric similarity, essential-operation checks, and feature-presence F1, without an explicit graph-matching term. The operation rewards may provide useful structural proxies, but they do not by themselves specify any graph-relevant reward. If such reward is not used, the claims should be revised accordingly. This ambiguity affects a central methodological contribution.

2. The alleviation of long-tailed operation distributions is not demonstrated.

Section 4 constructs complexity tiers using BRep face counts, while Appendix A.2 provides a qualitative gallery. Neither shows a clear distribution of CAD operations in the datasets. Greater face count does not necessarily imply greater coverage of rare operations. The paper does not provide operation-frequency histograms for the source data and TopoCAD-1M. Consequently, the claims of improved rare-operation coverage remains insufficiently supported.

3. The ablation does not establish the benefit of topology-guided initialization when ToRPO is used

Table 3 compares basic SFT, topology-guided SFT, and topology-guided SFT followed by ToRPO, but omits basic SFT + ToRPO without topology-guided SFT. As written, Eq. (6) evaluates generated programs and does not appear to require a topology-guided SFT initialization. The missing comparison is therefore necessary to assess whether topology-guided SFT provides additional value after optimization.

4. Baseline training-data fairness and the meaning of "cross-dataset" evaluation are unclear.

TopoCAD-1M incorporates ABC, DeepCAD, and a CC3D subset, while Table 2 evaluates on DeepCAD and CC3D, among other benchmarks. The paper does not adequately specify whether each baseline result comes from retraining or previously reported numbers.

Separately, in line 377, "Table 2 reports quantitative results of cross-dataset for point-cloud conditioned CAD reconstruction", evaluation on a dataset represented in the training corpus does not establish generalization to an unseen dataset. The "cross-dataset" label therefore needs qualification.

5. The evaluation does not directly measure the topology/structural-fidelity.

Section 4 and Appendix A.1 report CD, IoU, and IR. None directly measures agreement with the target constraint graph or recovery of its intended operations. This is consequential because the paper motivates its approach by arguing that geometric similarity alone is insufficient. These should be complemented by a direct structural measure, such as typed constraint precision/recall/F1. RessR_{\mathrm{ess}} and RfeatR_{\mathrm{feat}} should also be included in the metrics.

6. λcon\lambda_{\mathrm{con}} and λcode\lambda_{\mathrm{code}} in Eq. (5)

Eq. (5) introduces λcon\lambda_{\mathrm{con}} and λcode\lambda_{\mathrm{code}}, but their numerical values are not reported in the paper.

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

TopoGen studies executable CadQuery program generation from point clouds or multi-view images. It argues that geometric similarity alone does not establish structural fidelity: a generated program can approximate the target silhouette while omitting holes or replacing a design-defining operation with a simpler construction.

The method first extracts a constraint graph from ground-truth BReps using OCCT and a custom geometric engine. The graph includes geometric entities, directional and parametric relations, and incidence relations. Qwen3-VL-4B is fine-tuned to predict a serialized constraint description followed by the corresponding CadQuery program. The point-cloud pipeline also allocates 35% of its sampling budget near sharp edges. At inference, ground-truth constraints are not supplied.

The second stage, Topology Relative Policy Optimization (ToRPO), executes groups of sampled programs and optimizes a clipped, group-relative objective with KL regularization. The stated reward combines voxel IoU, essential-operation checks, feature-presence F1, and normalized Chamfer and Hausdorff terms. The authors also construct TopoCAD-1M by augmenting approximately one million CAD models with structural annotations and adjusting the face-count distribution. Experiments compare image- and point-cloud-conditioned reconstruction on four datasets and report a progressive ablation of the training pipeline.

Critical strengths
  1. Geometric similarity and program executability do not necessarily imply preservation of design-relevant features or structural relations. Moving beyond surface resemblance toward structural fidelity is a well-motivated objective.
  2. The framework provides a practical way to incorporate structural knowledge into program generation. Training the model to predict constraint descriptions alongside CadQuery programs provides a shared structural target for point-cloud and image inputs. Ground-truth constraint graphs are not required at inference time.
Critical weaknesses
  1. M3-to-M4 comparison adds the entire RL stage. Without a geometry-only RL baseline under matched training conditions, the gains cannot be attributed specifically to essential-operation and feature rewards rather than additional execution-based optimization.
  2. The manuscript itself describes prior work on visual-to-CAD program generation and execution-based optimization. ToRPO resembles a standard group-relative policy optimization objective, making the main contribution the structural supervision, reward design, and annotated dataset. Stronger controlled experiments are therefore needed to establish the value of these additions.
Questions for authors
  1. How accurately can TopoGen reconstruct CAD models when important details are occluded in the input images or the underlying structure is highly complex? Could the authors provide evaluations across different levels of visibility and structural complexity, including local-feature accuracy and failure cases? How does the method handle structures that are invisible in all input views?

  2. What is the complete pipeline for constructing TopoCAD-1M? Could the authors describe the specific steps or pipeline?

Comments on the paper

1 comment

Dear Reviewers, the authors have posted their comments in response to the reviews. Please make an effort to discuss your review and the authors' response. - regards AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.