I wonder if any of the LLMs are smart enough to verify their work the way humans model 3D objects, i.e. by superimposing the original design on top of their work and checking for differences.

This is actually something I'm working on as part of a more formalized version of this benchmark, both as a tool to assist LLMs in doing this (especially for ones with a more poor vision model), but also to properly measure how well they've done.

Still in progress, but early demo showing difference zones: https://non.io/video/fidelity-lens.mp4