acceptodds
Under review as a conference paper at ICLR 2027

ParaSketchFM: Joint Continuous–Discrete Flow Matching for Parametric CAD Sketch Generation

10 views

Abstract

Despite substantial progress in parametric CAD sketch generation, existing generators still exhibit primitive-type distribution discrepancies on SketchGraphs, leaving room to improve generation quality and distributional fidelity. We therefore introduce ParaSketchFM, a joint discrete–continuous flow-matching model with geometric regularization. The model couples absorbing categorical paths with continuous geometric flows over unordered primitive sets and incorporates a one-sided anchor-distance regularizer, addressing two key challenges: matching primitive-type distributions and mitigating geometric contraction during joint generation. Under a shared evaluation protocol on SketchGraphs, ParaSketchFM achieves an FID of 5.22, precision of 0.431, and recall of 0.490, with a maximum per-type slot-frequency deviation from the test set of 0.31 percentage points. These results establish a new performance benchmark under this protocol.

Primary area:
generative models
Keywords:
CAD sketch generation, flow matching, discrete–continuous generative models, parametric CAD, permutation-invariant set generation, geometric regularization, SketchGraphs

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

This paper proposes ParaSketchFM, a joint discrete-continuous Flow Matching framework for parametric CAD sketch generation. It uses a shared permutation-equivariant Transformer to jointly generate primitive types and continuous geometric parameters. The method further introduces type-probability gating and a one-sided anchor-distance regularizer to address primitive-type frequency discrepancies and geometric contraction. Experiments on SketchGraphs show improved FID and balanced precision and recall compared with several baselines.

Critical strengths

1.The paper addresses the joint generation of discrete primitive types and continuous geometric parameters, an important problem in parametric CAD sketch generation.

2.The proposed geometric regularizer explicitly addresses near-point contraction, and the paper provides ablation studies to examine its effect.

3.The experiments use a shared evaluation protocol and report multiple metrics covering generation quality, precision, recall, and primitive-type distributions.

Critical weaknesses
  1. The novelty over existing CAD sketch generation methods is limited. SketchDNN has already explored joint discrete-continuous generation for CAD sketches using fixed primitive slots and a permutation-equivariant architecture. ParaSketchFM introduces a different flow-matching formulation, type-probability gating, and geometric regularization, but the paper does not sufficiently establish the significance of these additions beyond the reported improvements. The authors should more clearly distinguish the proposed method from existing approaches and explain which limitations are specifically addressed by the new formulation.

  2. The problem formulation is relatively restrictive, and its practical significance is unclear. The method assumes a fixed set of 16 primitive slots, with unused slots represented by NONE. This is suitable for a bounded sketch-generation setting, but does not address more general structural decomposition problems, such as determining the number and arrangement of primitives needed to represent a target shape. The paper should better clarify the intended scope of the method and the practical problems it addresses beyond distributional sketch generation.

  3. The evaluation does not fully establish the practical usability of the generated CAD sketches. FID, precision, recall, and primitive-type frequencies characterize aspects of distributional quality, but do not directly establish whether generated sketches satisfy geometric constraints or are suitable for downstream parametric modeling. The paper also acknowledges that it does not generate geometric constraints. Additional evaluation of geometric validity and downstream usability would strengthen the practical claims.

Questions for authors

see weakness

Reviewer_nb7LOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper applies flow matching to 2D CAD sketches, mixing continuous and discrete parameters and applying geometric regularization. They improve on SketchDNN and focus on matching the distribution of primitive types

Critical strengths

The work shows real progress on a problem and demonstrates that a joint representation of discrete and continuous features, combined with flow matching, can improve the distribution of generated objects. Their FID showed significant improvement. The compare against multiple relevant prior methods and demonstrate improvements in multiple appropriately chosen metrics.

They identify a failure mode (contraction of point) and propose a solution to address it.

Critical weaknesses
  • I would like to better understand why it contracts geometry to zero. The pairwise regularization feels a little like a band-aid.
  • Unsure if the increased score and primitive distribution translates to actual better CAD sketches.
  • Unsure about the significance of unconditional of 2D CAD Sketch generation (most attention is on 3D CAD sketches)
  • Unsure if the output is fit for downstream CAD applications

** Post-Rebuttal** I appreciate the work they have done to address my concern; the reimainf concern is mainly significant, but I appreciate the contribution and I think I will stick with a rating of 'weak-accept'

Questions for authors
  • Why do you think flow matching produces better results than diffusion in this setting? Which aspects of your experiments help explain the improvement?
  • SketchDNN uses another method to prevent geometric parameters from decaying toward zero. You test this normalization on a model already trained with your anchor-distance regularizer. Do these mechanisms interact? Does their approach reduce contraction in JFM trained without the regularizer?
  • You mention that JFM achieves better FID than CFM but matches the primitive-type frequencies slightly less closely. Do you consider this difference to be within experimental variability, or is there a systematic explanation?
  • Can the method support conditional generation? Unconditional generation is the stated scope, but evidence of controllability would strengthen its significance.
Reviewer_TzgEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ParaSketchFM addresses unconditional generation of parametric CAD sketches through joint flow matching (JFM) of continuous geometry and discrete primitive types and construction flags. It introduces an anchor-distance regularizer to mitigate geometric shrinkage. Under the shared evaluation protocol, the regularized model achieves the lowest FID and highest recall among the evaluated methods, while substantially reducing sampling time relative to the retrained SketchDNN.

Critical strengths
  • The paper provides a useful comparison of continuous-only, discrete-only, and joint generation using a shared backbone and evaluation protocol.
  • The model achieves a clear improvement over the continuous-only baseline.
Critical weaknesses
  • The representation, backbone, and type-probability weighting closely follow SketchDNN, while joint continuous–discrete flow matching also has precedents. The main contribution appears to be an adaptation to CAD with additional geometric regularization.
  • Although the paper clearly reports this dependence, it remains unclear whether the regularizer addresses a broader difficulty in joint generation or compensates for the chosen gating scheme. Comparing different gates with and without regularization would help clarify this.
  • The improvements are mainly demonstrated through metrics computed from rendered images. These metrics provide limited evidence about geometric validity, endpoint connectivity, or degenerate primitives.
Questions for authors

Please see the weaknesses.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

We wanted to kindly check whether you have had a chance to review our responses to the reviewers’ comments. We would be grateful if you could let us know whether anything remains unclear. If possible, receiving your feedback before the discussion period ends would help us address any remaining concerns in a timely manner. We would be very pleased if the additional work we have provided addresses your concerns.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.