| 基于非自回归扩散模型的参数化CAD设计 |
投稿时间：2025-10-19 修订日期：2025-12-19  点此下载全文 |
| 引用本文： |
| 摘要点击次数: 3 |
| 全文下载次数: 0 |
|
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| 中文摘要:在计算机辅助设计（CAD）领域，提升参数化设计的自动化水平已成为关键任务。然而，现有的非自回归(Non-Autoregressive，NAR)模型虽然具备较高的生成速度，但在设计精度上仍有不足；另一方面，基于扩散的生成模型在此类特殊语法场景中面临收敛缓慢及结构保持困难等问题。为此，本文提出ParaCADGen，一种面向CADQuery代码生成的非自回归扩散模型。该模型在建模过程中引入了三项关键设计：1) 面向CADQuery API的专用分词策略，以增强语义及几何意图解析；2) 基于语法结构的感知掩码机制，在扩散迭代中优先优化核心语法片段；3) 结合结构约束的损失函数，以提升生成结果的几何一致性与语法完整性。实验结果表明，ParaCADGen的推理速度较主流自回归模型提升约6-10倍，生成质量优于现有非自回归模型，展现了其在参数化CAD设计任务中的潜在应用价值。 |
| 中文关键词:参数化CAD 非自回归模型 扩散 CADQuery 结构约束 |
| |
| Parametric CAD Design based on Non-Autoregressive Diffusion Model |
|
|
| Abstract:The automation of parametric design is a critical task in Computer-Aided Design (CAD). While current Non-Autoregressive (NAR) models offer high generation speed, they often lack design precision. Diffusion-based models, conversely, face slow convergence and struggle with structure preservation in specialized syntax scenarios. To this end, this paper proposes ParaCADGen, a non-autoregressive diffusion model for CADQuery code generation. The model incorporates three key designs: a specialized tokenization strategy for the CADQuery API, a syntax-structure-aware mask mechanism, and a structure-constrained loss function to improve geometric consistency and syntactic integrity. Experiments show that ParaCADGen increases inference speed by approximately 6 to 10 times over mainstream autoregressive models, and its generation quality surpasses existing non-autoregressive models. These results demonstrate its potential value for parametric CAD design tasks. |
| keywords:parametric CAD non-autoregressive model diffusion CADQuery structural constraint |
| 查看全文 查看/发表评论 下载pdf阅读器 |
次