It’s interesting, all these examples look exactly as what I remember a typical result at an intro graphics design would look like, in that the idea might be good, but it gets very sloppy (pun unintended) in the details, such as alignment, consistency – especially in the visual hierarchy, or typography.

What’s also interesting to me is that requesting ”professional typography” did not fix the problem at all. And I would guess that there is a staggering amount of examples in the training data.

Maybe inference is just wobbly by nature? The same as text output has the characteristic randomness, even design output cannot resist being – for a lack of a better word – _hand made_?