Comment by vunderba

> I'm honestly unsure what could be improved at this point.

That's because you're focusing a little bit too much on visual fidelity. It's still relatively trivial to create a moderately complex prompt and have it fail miserably.

Even SOTA models only scored a 12 out of 15 on my benchmarks, and that was without me deliberately trying to "flex" to break the model.

Here's one I just came up with:

  A Mercator projection of earth where the land/oceans are inverted. (aka land = ocean, and oceans = land)