The demand for Computer-Aided Design (CAD) in industrial product development continues to grow, motivating a shift from traditional graphical interfaces toward automated parametric modeling workflows that support rapid iteration and interactive exploration. In such settings, generation systems must produce executable and geometrically consistent parametric code while maintaining low inference latency. Although non-autoregressive (NAR) generation methods are attractive due to their parallel decoding efficiency, they often struggle with structural validity, geometric consistency, and semantic fidelity when applied to parametric modeling languages. To address these challenges, we propose Para-CADecoder (PCAD), a structure-aware non-autoregressive diffusion system for text-driven parametric CAD code generation, implemented using a script-based three-dimensional modeling library. PCAD integrates a CADQuery-oriented tokenizer that preserves function-level semantics, a structure-aware masking mechanism within the diffusion recovery process, and geometry-aware constraints designed to improve structural consistency and first-pass executability. Experimental results show that PCAD achieves inference speeds 4-8 faster than typical autoregressive models, while improving geometric fidelity and executability compared with existing non-autoregressive baselines. These results highlight a practical speed-quality trade-off and suggest that non-autoregressive diffusion models can effectively support interactive and early-stage parametric CAD ideation, where rapid feedback and editable code representations are essential.
Sun et al. (Wed,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: