Yeah. The newer versions of Dall-E, GPT-Image, and similar cloud apps seem to have subnetworks specifically for text. They're worlds better than they were just a year or two ago. Like, you can twist them into generating semi-coherent-looking text logos, or make a cartoon character wearing a T-shirt with some text on it.

I've seen some complex pipelines with open models, where people train loras specifically to fix the things the base models suck at, like hands, text, etc.

But it's still a really dumb idea to generate a whole presentation slide or infographic that way, for a wide variety of reasons. If you ever get decent results, just consider yourself lucky. I mean, even the people with the skills to do this well (who are few and far between) would find it way more trouble than just, you know, making slides the normal way.

The incompetence we keep seeing from Microsoft is staggering. I can only assume this is malicious compliance. I imagine some exec said "everyone needs to use AI for everything" and everyone below them said "okay you dumb fuck, here you go".