GenCAD

Original Article ↗ Hacker News Discussion ↗

What GenCAD Produces

  • Discussion centers on the claim that GenCAD “converts CAD latents into parametric CAD commands” and “generates the entire CAD program.”
  • Output is clarified as DeepCAD-style JSON: a sequence of sketch/extrude (pad) operations derived from Onshape data, i.e., a feature history, not a mesh.
  • This history is conceptually CAD-agnostic but in practice still kernel/application dependent, and does not currently map cleanly into arbitrary CAD tools or editable histories elsewhere.

CAD Technology Context

  • Several comments explain that real CAD behavior depends heavily on the geometry kernel and tolerances, especially for fillets, blends, and barely-intersecting surfaces.
  • Portable formats like STEP typically lose the operation history for this reason.
  • GenCAD is described as operating at a CSG-like abstraction (sketch + extrude), with B-rep used only as a downstream representation.
  • It currently supports only simple 2D sketches (lines/arcs/circles) and extrusions; no revolve, fillet, chamfer, drafts, or complex surface workflows.

Perceived Utility and Limitations

  • Many see the demonstrated examples as extremely basic (often a single extrude) and far from “real” parametric CAD work.
  • Several argue that the hard part of CAD is constraints, dimensions, tolerances, and editability; GenCAD does not yet address these.
  • Some view it as a solution in search of a problem; others say they personally struggle with CAD and would welcome reliable sketch/image → parametric model tools.

Practical Usability and Training Constraints

  • Attempts to run the Docker setup exposed missing dependencies; the containerization is criticized as brittle.
  • A user reports that on non-training images, even simple ones, the model almost never produces correct output.
  • The paper’s own limitations section (paraphrased in the thread) says it is trained mostly on noise-free, isometric CAD renders in a very specific visual style and with a very restricted operation vocabulary, which explains poor generalization.

AI/LLM Integration and Alternatives

  • Multiple commenters discuss combining GenCAD-like models with multimodal LLMs: text → image → CAD, or CAD-as-code workflows.
  • There is extensive mention of using LLMs today with OpenSCAD, CadQuery, Build123d, or custom languages; experiences range from “works great for simple parts” to “painful and brittle.”
  • Other AI‑CAD efforts and open-source kernels are cited, plus a recent survey suggesting the field is moving quickly beyond this work.

Meta and Miscellaneous Points

  • Some emphasize that geometric kernels are intrinsically hard; in comparison, CAM toolpath generation is “just” optimization once good geometry exists.
  • Minor side discussions touch on font licensing, autoplay video on the site, mobile layout issues, and containerization (Docker vs Nix).