acceptodds
Under review as a conference paper at ICLR 2027

SPARC: Compiling Visual Assembly Intent into CAD Assemblies

Abstract

Automatically assembling a set of existing CAD parts from a single reference image requires both visual reasoning and geometric verification. A vision-language model can infer part roles, relative layouts, and likely connections, but these predictions may not be supported by the actual part geometry. Conversely, B-Rep geometry can verify and realize assembly relations, yet often admits multiple plausible interface assignments. We present SPARC, a framework that requires no task-specific model training and formulates reference-guided CAD assembly as a compilation process from visual assembly intent to executable CAD constraints. Given an unordered set of B-Rep parts and a reference image, SPARC first generates a structured visual assembly proposal. Relation-specific geometric routines resolve feature references and construct candidate assembly plans, while assembly-level selection compares these plans using geometric evidence and cross-part consistency checks. Pre-execution preparation may further revise relations and interface bindings before compiling the resulting state into native CAD constraints. FreeCAD executes the program and returns execution and interference diagnostics for bounded local repair. We further introduce SPARC-Bench, a benchmark of 310 assemblies constructed from two public CAD assembly datasets. Experiments evaluate visual consistency, executed assembly-graph validity, interference, and end-to-end assembly success. SPARC achieves a 94.84% end-to-end assembly success rate on SPARC-Bench. Across the evaluated interface-ambiguity groups, the benefit of assembly-level plan replacement grows with ambiguity, supporting the need to resolve local geometric choices in the context of the complete assembly.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
CAD Assembly, Vision-Language Models, Multimodal Reasoning, Geometric Reasoning, B-Rep, Constraint-Based Assembly

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_Ru1jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SPARC, a framework that formulates the assembly of existing B-Rep parts from a single reference image as compilation from visual intent into executable CAD constraints. Without task-specific model training, the system combines VLM-generated assembly proposals, geometric interface grounding, assembly-level plan selection, and local repair guided by FreeCAD execution feedback. The authors introduce SPARC-Bench, comprising 310 assemblies, and report a 94.84% end-to-end assembly success rate. Validation includes comparisons with existing methods, component ablations, and an analysis of performance across different levels of interface ambiguity.

Critical strengths
  1. Rather than merely predicting part poses, SPARC grounds visual relations in existing B-Rep interfaces and compiles them into native CAD constraints, using execution diagnostics to repair errors. This produces executable outputs and directly addresses the central difficulty that individually compatible interface assignments can conflict at the assembly level. The alignment between the problem formulation and the system design is a meaningful strength.
  2. Table 2 shows that disabling assembly-level plan replacement or CAD-feedback repair reduces the success rate from 94.84% to 81.61% and 87.10%, respectively. Moreover, Table 4 shows that the benefit of plan replacement increases with interface ambiguity, consistent with the paper’s central motivation. These results provide stronger support for the design than the final system’s performance against external baselines alone.
Critical weaknesses
  1. Both Visual Assembly Consistency and the topology-consistency component of Executed Assembly-Graph Validity rely on majority voting by three VLMs, with some of these models also used as planning backbones. The paper does not report expert human validation or inter-evaluator agreement, leaving the accuracy of these judgments and their potential correlated biases insufficiently characterized. Successful CAD solving and interference checks provide objective evidence, but do not independently establish that the intended assembly has been correctly realized. This limits confidence in the reported 94.84% success rate.
  2. In Table 3, Local Greedy and Static Beam retain the initial relations and disable post-execution repair, whereas full SPARC additionally permits relation revision and feedback-driven repair. The authors explicitly acknowledge that this compares complete configurations, but the experiment cannot determine how much of the performance gap comes from the selection mechanism rather than these additional capabilities or execution costs. Table 2 establishes that the module benefits the full system, but does not replace a capability- and budget-matched comparison against alternative search methods. This limits the evidence for the core algorithmic contribution.
Questions for authors
  1. Have the VLM judgments been validated against expert human assessments? What is the agreement among the three evaluators, and do the success rates or method rankings change when the planning model is excluded from evaluation?
  2. Does SPARC retain its advantage over Static Beam when relation-revision capabilities, pre-execution preparation, feedback repair, and CAD-call budgets are matched?
  3. Can the authors provide sensitivity analyses for the manually specified scoring weights and replacement margins in Appendix A.3, together with results stratified by assembly family and part count, to establish whether the gains extend beyond a small set of predefined structures?
Reviewer_QBShOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SPARC proposes a training-free framework for assembling existing CAD parts from a reference image. A VLM generates an initial assembly plan, followed by geometric interface matching, assembly-level plan selection, CAD constraint execution, and local repair. The authors introduce a 310-case benchmark and report a 94.84% end-to-end success rate.

Critical strengths
  1. Complete executable pipeline. The framework integrates visual planning, geometric verification, and native CAD execution into a working assembly system.
  2. Strong benchmark performance. The reported success rate and ablation results demonstrate the practical value of assembly-level selection and execution feedback.
Critical weaknesses
  1. Limited benchmark generalizability. The evaluation contains only 310 assemblies, mostly small and relatively simple cases from public datasets. Generalization to complex, real-world assemblies remains unclear. Potential overlap with VLM training data is also not investigated.
  2. Limited methodological novelty. The VLM primarily generates an initial plan, while subsequent decisions rely heavily on handcrafted geometric routines, scoring functions, and repair rules. The framework does not demonstrate substantial integration of LLM reasoning with geometric tools and execution feedback.
Questions for authors
  1. Can the authors evaluate SPARC on larger, independently created real-world assemblies not drawn from existing public benchmarks?
  2. How do the authors assess potential training-data contamination from the public CAD datasets?
  3. What is the contribution of VLM-driven reasoning beyond initial plan generation? Could the authors clarify or quantify its role in geometric decision-making and feedback-driven repair?
Reviewer_5wAEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

SPARC addresses assembly reconstruction from a reference image and a fixed set of CAD parts. A VLM proposes an assembly plan and geometric hints, which are processed by a series of search and geometric operations to find the final CAD programs that produce the target shape. The paper introduces SPARC-Bench, containing 310 assemblies from AutoMate and Fusion 360 Gallery, and reports high assembly success without task-specific model training.

Critical strengths
  • The paper builds an integrated system that predicts executable CAD constraints given a reference image. It reports strong performance on its benchmark.
  • The idea of using a VLM to infer high-level part relationships makes sense.
Critical weaknesses
  1. Questionable Evaluation Protocols. First, VAC and the topology component of EAGV rely on VLM judgments without independent human or geometric validation. It is very hard to calibrate these scores. The most important metrics regarding geometric accuracy metrics used in prior work are absent such as chamfer distance. The introduction of a brand new set of metrics is unjustified. This makes the evaluation very uninformative.

  2. Generalization of the heuristic system is not demonstrated. SPARC relies on extensive heuristic rules, manually specified scoring functions with lots of hyperparameters. Yet the number and diversity of assemblies used to develop these mechanisms and prompts are unknown. Performance is evaluated on only 310 selected assemblies, while the original datasets contain thousands. The eligible pool after filtering and the reason for this evaluation size are not reported. Since no task-specific training is required, a conventional training/test split does not explain the limit.

Questions for authors
  1. What is the reason for not using previous metrics like chamfer distance and part accuracy?
  2. How many examples are used during heuristics/prompt development? How did the authors determine extensive hyperparameters (e.g., the equation at line 782)
  3. For how many assemblies passed each filter, and why was evaluation limited to 310? Can the frozen system be evaluated on a larger independent sample?
  4. For JoinABLe experiments, why average candidate probabilities across five checkpoints? I think the checkpoints were just using different seeds and were supposed to measure sensitivity to seeds.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.