Source-linked AI summary
PrimitiveCAD: An LLM-Based Point-to-CAD Reconstruction with Primitive-Aware Tokenization and Operation Alignment
Jian Gao, Kailin Bi, Jiamin Xu, Jinlan Xu, Gang Xu
TL;DR
Existing point-to-CAD methods often treat point clouds and CAD sequences generically, omitting CAD-specific primitive tokenization and supervision. PrimitiveCAD combines primitive-aware tokenization with operation-aligned SFT and feature-line-guided RL, and reports state-of-the-art CAD sequence reconstruction across public benchmarks.
Problem
Existing methods lack CAD-specific point-cloud feature extraction, CAD-adapted SFT supervision, and rewards for point-line-plane structure.
Source: §1 · p. 2Method
PrimitiveCAD uses primitive-aware point-cloud tokenization, an operation alignment loss during LLM SFT, and a feature-line alignment reward during RL.
Source: §1 · p. 2Results
PrimitiveCAD achieves state-of-the-art performance in CAD sequence reconstruction across multiple public benchmarks.
Source: §1 · p. 2Takeaways & Limitations
The framework integrates CAD structural information across point-cloud tokenization, SFT, and RL.
Source: §1 · p. 2The framework focuses on component-level generation and does not yet support complex multimodule assembly models.
Source: §5 · p. 8
Abstract
from arXiv · showhide
Large-model-based point-to-CAD generation holds immense potential for advancing industrial design and enhancing 3D modeling efficiency. However, most existing methods approach the problem as a general point-cloud encoding and token prediction task, neglecting the tokenization and supervision specifically for CAD-related primitives. As a result, these methods often struggle to accurately reconstruct the intricate primitive structures. To address this limitation, we propose PrimitiveCAD, a novel multi-stage paradigm for point-to-CAD reconstruction that enhances the geometric accuracy of generated CAD models while better preserving critical geometric features. First, we introduce a primitive-aware point cloud tokenization model, enabling the system to learn more robust geometric representations from CAD point clouds. Next, we perform supervised finetuning on a large language model (LLM) and introduce an operation alignment loss to align key CAD operation frequencies, thereby improving the preservation of global shape features. Finally, we incorporate reinforcement learning (RL) and introduce a feature-line alignment reward to further reduce stochasticity and enhance the fine-grained preservation of geometric features. Experiments on the DeepCAD and Fusion360 datasets show that our method achieves state-of-the-art performance in code validity, geometric accuracy, and geometric feature preservation.
1 Introduction
Existing point-to-CAD methods lack CAD-specific point-cloud features, sequence supervision, and structural rewards, motivating PrimitiveCAD’s CAD-aware training stages. The framework reports state-of-the-art CAD sequence reconstruction across public benchmarks.
1 Introduction
- Existing methods lack CAD-specific point-cloud features, CAD-adapted SFT supervision, and rewards for point-line-plane structures.Source: §1 · p. 2
- Primitive-aware PTv3 fine-tuning supplies discriminative CAD-point features, while operation alignment constrains key-operation frequencies during SFT.The authors report SFT alone performs on par with existing baselines that incorporate RL.Source: §1 · p. 2
- The RL feature-line reward explicitly matches topological features to improve fine-grained reconstruction and reduce stochasticity.Traditional volume-based IoU can miss fine details and provide insufficient guidance when small features are absent.Source: §1 · p. 2
- The framework combines primitive-aware point-cloud tokenization, LLM sequence generation through SFT and RL, operation alignment loss, and feature-line alignment reward.The operation loss aligns key-operator frequencies; the reward addresses limitations of volume-based IoU for small features and complex topologies.Source: §1 · p. 2
- PrimitiveCAD achieves state-of-the-art performance in CAD sequence reconstruction across multiple public benchmarks.Source: §1 · p. 2
2 Related work
Prior point-to-CAD work spans geometric fitting, symbolic sequences, and LLM-generated CAD programs, but the cited approaches leave CAD-specific geometric structure insufficiently represented. PrimitiveCAD combines primitive-aware tokenization with operation-aligned SFT and feature-line-guided RL.
2 Related work
- CAD generation shifted from discrete operation sequences toward executable Python programs that produce parametric B-Reps.Source: §2 · p. 2
- LLM CAD approaches support text, image, and point-cloud inputs, yet token-level cross-entropy does not account for CAD-operation characteristics.Source: §2 · p. 2
- Prior reinforcement-learning approaches target stochasticity and output quality, while the cited work identifies fine-grained CAD feature alignment as unaddressed.Source: §2 · p. 2
- PrimitiveCAD’s pipeline assigns primitive-aware tokenization to the input, operation-frequency alignment to SFT, and feature-line alignment to RL.The feature-line reward covers lines, arcs, and circles for local geometric fidelity.Source: Figure 2 · p. 3, §2 · p. 3
- Earlier point-to-CAD methods used fitting or learned point-cloud representations, but often lacked primitive-aware geometric priors.Source: §2 · p. 3
3 Proposed Method
PrimitiveCAD converts point clouds into executable CAD scripts through primitive-aware tokenization, operation-aligned supervised fine-tuning, and reinforcement learning with geometric feedback. Its stages target both global construction logic and local geometric features.
3.1 Overview
- The method generates executable Python CAD scripts from point clouds, then refines them with supervised fine-tuning and reinforcement learning to reduce stochasticity.The target scripts generate parametric B-Reps.Source: §3.1 · p. 3
3.2 Primitive-Aware Point Cloud Tokenization for CAD Models
- PTv3 fine-tuned on normal estimation and primitive-instance segmentation supplies geometric features and primitive-parsing capabilities for later CAD code generation.The predicted point-wise outputs include normals and primitive-instance categories.Source: §3.2 · pp. 3–4
3.3 SFT with Operation Alignment
- Operation alignment regularizes global operator composition, complementing token-level code generation so outputs target both syntactic validity and CAD construction logic.The loss is designed to address missing or redundant operators, which can cause significant deviations in final geometry.Source: §3.3 · pp. 4–5, §3.3.1 · p. 5
3.4 RL with Feature-Line Alignment
- RL executes sampled code and rewards volumetric consistency and local feature fidelity, with feature-line matching across straight lines, arcs, and circles.The RL stage does not require complete CAD sequence labels and can use datasets without sequence annotations.Source: §3.4 · pp. 5–6
4 Experiments
Experiments compare PrimitiveCAD on DeepCAD and Fusion360, then ablate its SFT and RL components. The reported results favor the full approach, while component studies examine geometric accuracy, feature preservation, and generation robustness.
4 Experiments
- SFT training uses DeepCAD and CAD-Recode models, while Fusion360 guides RL; evaluation reports IoU, invalidity rate, and mean and median Chamfer distance.Quantitative comparisons use single-pass evaluation, while qualitative results show each method’s best of 10 candidates and the RL ablation also reports worst-of-10 results.Source: §4 · p. 6
4.1 Evaluation of SFT and RL
- On CADEvolve, the framework reconstructs models with fillet, chamfer, and shell operations while preserving geometric accuracy.The results are reported in the supplementary material.Source: §4.1 · p. 6
- PrimitiveCAD achieves the best overall performance on DeepCAD and Fusion360, with higher IoU and lower Mean CD and median CD than Cadrille using the same backbone.Training on the larger CAD-Recode dataset further enhances performance; Table 1 results exclude RL finetuning and test-time sampling.Source: §4.1 · p. 6, Table 1 · p. 7
- In best-of-10 qualitative comparisons, PrimitiveCAD better preserves fine details such as holes and right angles than CAD-Recode and Cadrille.The reported qualitative gains are attributed to primitive-label PTv3 finetuning, SFT operation alignment, and the RL feature-line reward.Source: §4.1 · p. 6, Figure 7 · p. 7
- On Fusion360, PrimitiveCAD without RL matches RL-finetuned Cadrille, and RL finetuning then outperforms Cadrille across all evaluation metrics.The comparison evaluates SFT-only and RL-finetuned PrimitiveCAD against Cadrille; Table 2 reports these variants.Source: §4.1 · p. 6, Table 2 · p. 8
4.2 Ablations
- PTv3 finetuning with primitive labels and the operation alignment loss jointly improve IoU and CD in the SFT ablation.The qualitative comparisons are consistent with the quantitative gains; Fig. 8 compares variants with and without PTv3 finetuning and operation alignment.Source: §4.2 · p. 6, Table 3 · p. 8, Figure 8 · p. 8
- Using 512 input points performs worst on IoU and CD, while 1024, 2048, and 4096 points perform comparably; 1024 provides a balance with input complexity.The 1024-point variant has the best IoU, while 4096 points yield slightly better CD; gains beyond 1024 are marginal.Source: §4.2 · pp. 7–8, Table 4 · p. 8
- RL improves robustness over SFT-only, and removing the feature-line reward lowers IoU and raises CD; under the same reward design, Dr. CPPO performs worse than the full model.The full model’s best–worst gaps are mean CD 0.59, median CD 0.037, IoU 7.3, and IR 0.87; Table 5 and Fig. 9 report best- and worst-of-10 results.Source: Table 5 · p. 8, §4.2 · p. 8, Figure 9 · p. 8
5 Conclusion
PrimitiveCAD generates CAD sequences from point clouds through a multi-stage approach, with primitive-aware representations and aligned SFT and RL objectives. The current framework focuses on component-level generation and does not yet support complex multimodule assemblies.
5 Conclusion
- PrimitiveCAD generates CAD sequences from point clouds through primitive-aware representations followed by LLM-based SFT and RL.The conclusion associates PTv3 fine-tuning with primitive labels with geometric representations, operation alignment with global shape accuracy, and feature-line alignment with geometric detail.Source: §5 · p. 8
- Experiments demonstrate effectiveness in geometric accuracy, code validity, and reconstruction quality.Source: §5 · p. 8
- The current framework focuses on component-level generation and does not yet support complex multimodule assembly models.The authors identify such models as future work.Source: §5 · p. 8