Spectral Labs drops SGS-1 to turn sketches into parametric CAD files

Most 3D generators look impressive until an engineer actually opens the file. What you usually get is an uneditable polygon mesh—great for video games, completely useless for manufacturing.

Spectral Labs just released SGS-1, a foundation model built specifically for structured, parametric CAD geometry that can actually be sent to a factory floor.

Real CAD parts, not hollow meshes

Existing image-to-3D models spit out “dumb” meshes. If you need to tweak a hole diameter or adjust a flange by two millimeters in SolidWorks, you are essentially out of luck.

SGS-1 generates native B-Rep geometry and exports it directly as editable STEP files. Users can feed it a simple reference photo, a rough hand-drawn napkin sketch, an engineering drawing, or even an uneditable STL scan.

The model converts these inputs into parametric solids that slot straight into standard engineering software. In one demo, researchers fed SGS-1 a partial render of a roller assembly and asked for a mounting bracket; the model generated a fitting bracket directly within the assembly context.

Benchmarking spatial reasoning

To see how it stacks up, the team tested SGS-1 against HoLa, an open-source image-to-CAD model, and OpenAI's GPT-5 thinking, which was tasked with writing CadQuery code.

Using a benchmark of 75 medium-to-high complexity mechanical parts, each model got 10 attempts to produce a valid, watertight solid. According to the team, GPT-5 struggled heavily with spatial awareness, often producing geometry that was either broken or far too simplistic.

SGS-1 consistently produced viable geometry across the test set, though it still has clear limitations. The team notes it cannot handle complex organic curvature, struggles with razor-thin walls due to 3D resolution limits, and cannot generate multi-part assemblies in a single pass yet.

Why it matters

For mechanical engineers and hardware startups, turning raw 3D scans or rough concepts into usable CAD files is hours of tedious reverse-engineering. If SGS-1 can reliably automate even half of that pipeline, it is worth testing immediately.

More broadly, this exposes a real blind spot for general-purpose frontier LLMs. Text-based reasoning models, even as advanced as GPT-5, simply do not understand physical 3D space well enough to write production-grade geometry from scratch.

We think targeting parametric B-Rep formats instead of game assets is the exact right move. Physical manufacturing runs on tight tolerances and STEP files, not polygons. Spectral Labs is tackling the boring, unsexy part of 3D, which is precisely where the real commercial value lives.