acceptodds
Withdrawn from ICLR 2027

PartCoder: Reconstructing Multi-File CAD Code Assemblies from Point Clouds

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

14 views

Abstract

Recent LLM-based CAD reverse-engineering methods recover editable CAD code from point clouds but represent entire objects as monolithic programs. This creates generation and representation bottlenecks because the model must reconstruct the complete geometry within a single script, while the output lacks the explicit part structure and part-level controllability of industrial CAD assemblies. We introduce PartCoder, the first method for reconstructing multi-file CAD code assemblies directly from point clouds. PartCoder hierarchically decomposes a dense input point cloud into geometrically coherent parts, completes newly created interfaces, reconstructs each part using a generative CAD code backbone, and combines the components through a central assembly file. It retains only decompositions that improve the overall reconstruction, adaptively selecting the number of parts and enabling existing backbones to handle more complex objects without retraining. We introduce PrIMe3D, a benchmark of 1,000 mechanical objects generated through a prompt-to-image-to-mesh pipeline for evaluating the emerging task of converting AI-generated 3D assets into editable CAD models. Averaged across three backbones, PartCoder improves mean IoU over the corresponding monolithic baselines by 1.1, 7.7, and 11.1 percentage points on Fusion360, OmniCAD, and PrIMe3D, respectively, with larger gains on more complex datasets while producing modular CAD assemblies with independently editable parts.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
CAD, CAD reverse engineering, CAD code generation, multi-file CAD assemblies, point clouds, large language models for CAD

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_34bqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces PartCoder, a framework for reconstructing multi-file CAD code assemblies directly from point clouds. It uses PartField and hierarchical clustering to progressively decompose the input geometry, completes missing interfaces introduced by decomposition, and reconstructs each part independently using an existing CAD generation backbone. The resulting parts are combined through a central assembly file. Chamfer distance between the assembled reconstruction and the input point cloud determines whether a split is accepted and decomposition continues, without additional backbone training. The authors also introduce PrIMe3D, a benchmark for converting AI-generated 3D assets into CAD models. Experiments with three generation backbones across three datasets show improvements in geometric reconstruction quality, with larger gains on more challenging datasets.

Critical strengths
  1. Practically relevant framework. The paper introduces PartCoder, a framework for reconstructing multi-file CAD code assemblies directly from point clouds. By hierarchically decomposing a dense input point cloud into geometrically coherent parts and reconstructing them independently, PartCoder enables existing CAD generation backbones to handle more complex objects without additional training.

  2. Consistent empirical improvements. Experiments across three backbones provide evidence for the effectiveness of PartCoder. Averaged across these backbones, the reported mean IoU improvements over the corresponding monolithic baselines are 7.7 percentage points on OmniCAD and 11.1 percentage points on PrIMe3D. These results support the value of decomposing difficult reconstruction tasks into simpler part-level problems.

Critical weaknesses
  1. Insufficient ablation of the main components. The experiments do not adequately isolate the contributions of the method’s key components. The maximum-part-count ablation does not establish the individual importance of PartField-based decomposition, interface completion or adaptive stopping. Comparisons with simple spatial clustering or planar partitioning, reconstruction without interface completion and fixed decomposition granularity would help identify which design choices account for the observed gains.

  2. Limited validation of structural quality and practical editability. Although the decomposed parts need not correspond to the original manufacturing components, the current evaluation primarily measures geometric similarity and does not directly assess the quality of the modular organization or its usefulness for editing. Concrete part-level editing examples or evaluations, including modification, removal and repositioning, would strengthen the claims of controllability. An analysis of unintended overlaps, gaps, and interface misalignment would also help establish the engineering applicability of the resulting assemblies.

Questions for authors
  1. Contribution of the decomposition method: How much improvement does PartField-based decomposition provide over simpler alternatives, such as spatial clustering or planar partitioning? Could the authors provide controlled comparisons to establish the importance of the chosen decomposition method?

  2. Contribution of interface completion: How do reconstruction quality and code validity change when interface completion is removed? What are the typical failure cases of the current plane-and-convex-hull approximation for curved, non-convex, or perforated interfaces?

  3. Contribution of adaptive stopping: What advantages does adaptive stopping offer over reconstruction with a fixed number of parts in terms of quality and efficiency? Could continuing to evaluate deeper decompositions after the first non-improving split yield better results?

  4. Validation of practical editability: Could the authors provide concrete examples or task-based evaluations of part modification, removal, and repositioning to demonstrate the practical value of the modular outputs? Could they also analyze unintended overlaps, gaps, and interface misalignment between parts?

Reviewer_cF1nOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PartCoder proposes to convert 3D point clouds into editable CAD assemblies by splitting objects into simpler parts, generating code for each, and keeping splits that improve reconstruction. Across three backbones, PartCoder improves mean IoU on Fusion360, OmniCAD, and PrIMe3D and shows greater benefits on complex geometry while producing independently editable parts.

Critical strengths
  1. PartCoder takes the part-level CAD output into consideration, enhancing the capability of existing models in reconstructing complex shapes as multi-file CAD models from point clouds. The method is easy to understand and easy to implement.

  2. The paper is well written, and PartCoder makes clear improvements in reconstruction performance on complex-shape benchmarks( OmniCAD and PrIMe3D).

Critical weaknesses
  1. PartCoder will stop the part split loop once the Chamfer Distance does not decrease. This protocol omits a circumstance where the whole shape reconstruction performance may get better after one more split.

  2. PartCoder combines existing off-the-shelf models with part decomposition and interface completion. According to Table 2 and Table 3, the average number of final CAD model parts is relatively small but the runtime increases substantially .

Questions for authors
  1. Considering PartCoder uses CAD-Recode as one of its CAD reconstruction backbones, what are the runtime differences between PartCoder-CAD-Recode and CAD-Recode?

  2. How does PartCoder handle the completed interface when a part with a completed interface is further split in a later step.

  3. Could you please provide more details about the PrIMe3D benchmark, including example prompts based on mechanical engineering vocabulary, the filtering and checking process( if any) and whether reference CAD files are available?

Reviewer_4T8hOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces PartCoder, a training-free method for reconstructing multi-file CAD code assemblies from point clouds. It decomposes the input point cloud into clusters using PartField. Then, each cluster is fitted into a CAD model using pretrained LLMs and then, all generated parts are combined into an assembly. The paper also introduces PrIMe3D, a benchamrk of 1000 AI-generated meshes and their corresponding CAD programs. PartCoder is evaluated against cadrille, CADReasoner, CAD-Recode and closed-source models on Fusion360, OmniCAD, and PrIMe3D. The evaluation includes geometric fidelity, validity ratio, runtime and qualitative comparisons.

Critical strengths
  • PartCoder is an interesting training-free method for point cloud to multi-file CAD code generation
Critical weaknesses
  • The evaluation is very limited given the claims of the paper. The paper claims to reconstruct multi-file CAD assemblies, but the evaluation is largely limited to single-part CAD objects. The paper doesn't evaluate against actual CAD assembly samples available in ABC dataset [1] or Fusion 360 Gallery Assembly [2].
  • The paper claims to improve the controllability of the generated CAD, but no evaluation backs this statement. Moreover, the part clustering doesn't look very accurate in Figure 2.
  • The improvement over the baselines is marginal, and the largest gap is on the paper's own benchmark, PrIMe3D, which looks a lot more prismatic than the other two benchmarks.
  • The planar assumption is very strong and not backed by any real-world examples where that is the case. Most assemblies have interacting mates that make planar part interfaces unlikely.

[1] Koch, Sebastian, et al. "Abc: A big cad model dataset for geometric deep learning." 2019 IEEE/CVF conference on computer vision and pattern Recognition (CVPR). Ieee, 2019.

[2] Willis, Karl DD, et al. "Joinable: Learning bottom-up assembly of parametric cad joints." 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022.

Questions for authors
  • Could you evaluate your method against the baselines on CAD assembly samples from the ABC dataset [1] or the Fusion 360 Gallery Assembly [2]? This would back the SOTA claim in the multi-part CAD assembly setting.
  • Could you add an editability/controllability metric that showcases the claim that your hierarchical part-based pipeline that decomposes objects into modular CAD programs improves controllability.
  • Could you include the ground truth CAD assemblies for each sample in the PrIMe3D dataset, which would allow more precise evaluation instead of evaluating CD and IoU against the noisy meshes from Trellis.
  • Could you compare key metrics of the PrIMe3D dataset against the other benchmarks you evaluate your method on? Metrics like face count of the BReps per solid (associated with complexity), number of operations/single parts per assembly (if ground truth available).

[1] Koch, Sebastian, et al. "Abc: A big cad model dataset for geometric deep learning." 2019 IEEE/CVF conference on computer vision and pattern Recognition (CVPR). Ieee, 2019.

[2] Willis, Karl DD, et al. "Joinable: Learning bottom-up assembly of parametric cad joints." 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022.

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

Related papers

Discussion (0)

Sign in to comment.