Abstract
Computer-Aided Design (CAD) generative modeling is driving significantinnovations across industrial applications. Recent works have shown remarkableprogress in creating solid models from various inputs such as point clouds,meshes, and text descriptions. However, these methods fundamentally divergefrom traditional industrial workflows that begin with 2D engineering drawings.The automatic generation of parametric CAD models from these 2D vector drawingsremains underexplored despite being a critical step in engineering design. Toaddress this gap, our key insight is to reframe CAD generation as asequence-to-sequence learning problem where vector drawing primitives directlyinform the generation of parametric CAD operations, preserving geometricprecision and design intent throughout the transformation process. We proposeDrawing2CAD, a framework with three key technical components: anetwork-friendly vector primitive representation that preserves precisegeometric information, a dual-decoder transformer architecture that decouplescommand type and parameter generation while maintaining precise correspondence,and a soft target distribution loss function accommodating inherent flexibilityin CAD parameters. To train and evaluate Drawing2CAD, we create CAD-VGDrawing,a dataset of paired engineering drawings and parametric CAD models, and conductthorough experiments to demonstrate the effectiveness of our method. Code anddataset are available at https://github.com/lllssc/Drawing2CAD.