Input types we have not focused on yet

These input types are not off-limits. They simply are not what today's AI has been trained and optimised for, so results will look different from what you would get with a mechanical part or a photo. If you are curious, go for it — and tell us what came back.

Functional schematics and system diagrams

Block diagrams, wiring diagrams, system architecture layouts and P&IDs describe how components relate to each other, not the geometry of a single physical part. With no single shape to reconstruct, the output typically will not resemble a coherent manufacturable object.

Kinematic and mechanism diagrams

Textbook-style drawings of linkages, gear trains, cam mechanisms or multi-body systems (Geneva drives, four-bar linkages, walking-beam mechanisms). They look technical and often carry numeric labels, but those numbers are usually unitless link lengths or angles used for motion analysis, and the drawing shows several rigid bodies connected by joints rather than one manufacturable part.

This is one of the more interesting edge cases, because the individual components inside the diagram (a star wheel, a cam, a linkage arm) often *are* real mechanical shapes, just drawn schematically. If you upload one, say in the part brief what you are actually after:

"I want a 3D representation of the whole mechanism, to visualise how it moves"

"I'm only interested in the star wheel from this diagram — please reconstruct just that"

Without that context the system cannot know which of the two you mean, and today's AI is trained for the second case.

Electronic assemblies, modules and complete devices

PCBs, populated circuit boards, display or sensor modules, cable assemblies, connector pinouts or manufacturer spec sheets. Even with visible dimensions, the AI has not been trained to interpret electronic function as physical geometry, so the result tends not to reflect anything manufacturable.

Anything else you are curious about

Clothing, furniture, buildings, everyday objects — turning non-engineering objects into 3D models simply has not been our development focus so far. If enough of you want it, that could change.

A quick way to calibrate expectations

Ask yourself: is the value in the physical shape, or somewhere else (an electronic function, a system relationship, something non-mechanical)? The more the answer points to "the shape itself", the closer the input is to what today's model handles well.

Tried something unconventional? We want to know. Use the feedback button in the app and tell us what you tried and what came back — that is exactly the input that shapes what we train on next.