Concepts
These pages explain the ideas the LatticeLabs Toolkit is built on. They are conceptual background — orientation, not API docs — adapted from the project’s research notes.
A note on honesty: where these pages cite validity rates or benchmark numbers, those come from the published research literature (DeepCAD, BrepGen, Text2CAD, and others) and describe what the field has achieved. They are kept distinct from the toolkit’s own measured results, which are now real and reproducible — e.g. ll_brepnet segmentation at test mIoU 0.828, and ll_gen’s program-based generators producing measured-valid CAD (0.914 / 0.934, gated on real non-degenerate solids). Where a page shows a LatticeLabs number it is labeled as such; field numbers frame what is realistic, package pages report what this codebase outputs.
The pages
Section titled “The pages”- How geometry becomes tokens — why continuous 3D geometry has to be quantized into discrete tokens, and how geotoken does it adaptively.
- How neural networks generate CAD — the end-to-end pipeline from a prompt to a valid solid, and the two dominant strategies (token sequences vs. code generation).
- Inside CAD generation models — the architectures themselves: transformers, diffusion, VQ-VAE, and the GNN encoders that read B-Rep.
- The reality of AI CAD generation — what actually ships in production, why “propose → dispose” is the only reliable pattern, and why ll_gen is built that way.
The one idea to take away
Section titled “The one idea to take away”Neural networks output soft probability distributions over continuous spaces; CAD requires exact integer topology (face counts, edge connectivity) and precise parameters. Every technique below is, at bottom, a way to bridge that gap — by quantizing geometry into tokens, by encoding topology implicitly, or by making the network write code that a battle-tested CAD kernel executes.