My idea is simple.
Using machine learning it should be possible to generate a more realistic 3D model of a simplified 3D block object.
The blocks are just simple primitive 3D objects like different shapes of boxes and other primitives like planes, spheres and more.
The idea is that you put together the block objects into a crude model. Then you label it and let the machine learning software generate a more realistic model. This based on its previously trained samples and target models.
So if you build a house with these big blocks of shaped boxes then label it as a house or label it using a target image. The target image could be the inspiration data. The machine learning software will then generate a realistic looking house with much more 3D polygon detail.
Imagine something like a 3D game where you first build your world with big blocks and then let the software generate a realistic world from it. Pretty nice.