Skip to content

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.

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.