SwiftCAD: Efficient Parametric CAD Generation with Shared Decoder Transformers

Juyoung Kim, Seongjun Choi, Jiyeon Lim, Soomok Lee

Keywords: Computer-Aided Design (CAD), Parametric CAD Generation, 3D Reconstruction, Transformers, Shared Decoder, Efficient Architecture, Scalable Vector Graphics (SVG), Sequence-to-Sequence Learning

Abstract:

Maintaining precise geometric dimensions is critical in Computer-Aided Design (CAD). Recent works have utilized representations for 3D generation, such as voxels, point clouds, and polygon meshes. In parametric CAD generation, approaches using Scalable Vector Graphics (SVG) have demonstrated superior performance in preserving dimensions. However, industrial applications require efficient operation on limited hardware, while existing architectures incur computational costs due to redundant structural components.
In this paper, we propose two methods to improve efficiency. First, we remove redundant MLP layers to simplify the encoding process. Second, we adopt a weight-shared single-decoder (shared decoder) to jointly predict commands and parameters. To train and evaluate our methods, we use the CAD-VGDrawing dataset. Our method achieves comparable generation accuracy while reducing model size and inference time.

Submission Number: 13