IGeoCAD: Unlocking Intrinsic Geometric Awareness for CAD Command Sequence Generation

Chulong Zhong, Zhi-Yuan Zhang, Shuangping Huang, Yunqing Hu, Huiyuan Zhang, Tianshui Chen

Keywords: CAD Command Sequence Generation, Geometric Awareness, Sequence modeling

Abstract:

The CAD command sequence generation task aims to translate design requirements into an executable sequence of precisely parametric commands for 3D asset construction, facing the critical challenge of capturing complex geometric relationships among primitives, especially points and curves.
Previous methods attempt to enhance geometric awareness by differentiating commands from their parameters through parametric guidance.
However, this simple distinction only implicitly models geometric awareness, failing to adequately represent and enforce crucial relationships at both the point and curve levels.
In this paper, we contend that robust geometric reasoning requires a model to possess a dual awareness: point awareness to align with high-level design requirements, and curve awareness to maintain procedural consistency throughout the generation process. To this end, we introduce IGeoCAD, a framework that implements this dual awareness by explicitly modeling constraints through two simple yet effective modules: the Point Constraint-Aware Module (PCM) and the Curve Constraint-Aware Head (CCH).
Specifically, at the point level, PCM interprets the geometry of each command, aligning high-level requirements with keypoints and modeling spatial dependencies through coupled attention, thereby ensuring precise localization and consistent point constraints.
Concurrently, at the curve level, CCH fosters sequence geometric consistency by tracking dependencies across commands through a multi-task objective where an auxiliary head infers constraints to condition parameter prediction.
Collectively, these modules significantly enhance IGeoCAD's geometric awareness, enabling the generation of complex and precise CAD programs. Extensive experiments demonstrate the effectiveness of the proposed approach.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 6050