GenCAD turns images into editable parametric CAD programs, not just meshes

GenCAD is an image-conditional generative model that produces parametric CAD command sequences from a single rendering, yielding editable engineering models rather than the meshes, voxels, or point clouds typical of prior image-to-3D systems. The output is a full CAD program that a geometry kernel can replay into a B-rep solid, preserving the modifiability designers need for downstream manufacturing and design exploration.

The architecture stitches together four components: an autoregressive transformer encoder that learns latent representations of CAD command sequences, a contrastive model that aligns those latents with image embeddings, a latent diffusion model that samples command-sequence latents conditioned on an input image, and a decoder that turns the latents back into parametric commands. Because diffusion runs in latent space, the system can generate multiple distinct CAD candidates for the same image and supports retrieval against a library of roughly 7,000 existing programs.

The significance is practical: most image-to-3D work has prioritized visual fidelity over engineering usability, and a model that emits the construction history rather than a frozen mesh is far more useful for CAD workflows where parts must be parametrically tweaked, manufactured, or fed into further automated design pipelines.