There's probably a project to make here. Take those pixel2pixel models that are able to generate images from drawings, generalize them to 3D, say a pixel2vertex. Would be pretty neat.
The key would be to find a way to constrain the output space so that it's not so vast that it won't be possible to converge to something useful.
That would be very interesting, reminds me of an demo a couple years back of sketching buildings and and immediately being translated using procedural shape grammars into 3d.
I thought it was going to be a cool demo on changing 2D drawings to 3D objects.