Researchers introduced GenCAD, an image‑conditional system that produces full parametric CAD programs instead of just a solid mesh. By ingesting a reference photograph, the model emits a sequence of CAD commands that a geometry kernel can render into an exact B‑rep model. This approach preserves the editability engineers require for downstream manufacturing and design exploration in real‑world pipelines worldwide today.

At its core, GenCAD stitches together four stages. An autoregressive transformer encoder learns latent representations of CAD command sequences, while a contrastive model aligns those latents with image features. A latent diffusion model then generates command latents conditioned on the input picture, and a decoder translates them back into executable CAD code for downstream analysis and rapid prototyping in industry.

Because GenCAD outputs the full command history, designers can modify dimensions, constraints, or feature order without rebuilding the model from scratch. This fidelity bridges the gap between visual inspiration and manufacturable geometry, promising tighter integration of AI into CAD workflows. Early experiments show the system matches mesh‑based generators in visual quality while retaining true parametric control.