The ability to write text

[cmp-nct]

I've been looking at the examples from Nvidia and they claim their model can write text. the images are impressive.

Stable Diffusion is also able to write text similar to it, just that it is not readable.
I wondered a bit how they might have done it and the most likely approach is training of words instead of letters.
https://deepimagination.cc/eDiffi/

My guess is that they created images with popular words (like the top 1k words ?) maybe also short popular word groups.
Then they trained it. It allows for perfect tagging as well.
"House" could be tagged as @@house@@, so as to not destroy the actual "houses" it learned.

So their model likely does not know how to write, it just knows how to write certain words and groups of words. They likely look for "XXXX" inside the quotes and add the special markers.
Maybe custom training could do that ?