Beyond Meshes: Generating CAD Programs

Most AI-driven 3D generation outputs meshes, voxels, or point clouds — formats that look good but lack engineering precision and editability. GenCAD takes a different approach: given an image of a 3D object, it generates the complete parametric CAD command sequence needed to recreate and modify that design.

Architecture

Autoregressive Transformer Encoder: Learns latent representations of CAD command sequences
Contrastive Learning Framework: Bridges the gap between CAD command representations and CAD images via joint embedding
Latent Diffusion Model: Generates CAD command representations conditioned on input images
Decoder: Converts latent representations back into executable parametric CAD commands

Key Capabilities

Image-to-CAD generation: Produces both 3D solid models and full CAD programs from image renderings.

Design diversity: Generates multiple valid CAD outputs for the same image input, enabling design space exploration.

CAD retrieval: Searches a database of ~7,000 CAD programs to identify semantically similar designs from images.

Access and Significance

Code is available on GitHub (ferdous-alam/GenCAD), the paper is on arXiv (2409.16294), and an interactive demo is hosted on the project website.

Generating editable parametric CAD — rather than static meshes — is the step that makes AI design tools genuinely useful for manufacturing, architecture, and product design. If AI can produce a modifiable CAD history from a sketch or photo, it fundamentally changes the early-stage design workflow for engineers.

#gencad

GenCAD: AI System That Generates Editable Parametric CAD Programs from Images