It's also telling that these models have been trained on classical paintings in the public domain. They have the quiet stoicism of martyrs
also the AI is really into flowing loincloths and drapery...
Again though we can see that these systems have no internal representation of what they produce.
For instance even a beginning artist won't sometimes get confused about how many belly buttons a human body is supposed to have...everybody's internal model of how a human body is built requires exactly one, and you won't get two or three in some percent of cases.
(While of course an advanced artist's internal model of the human body will include much more anatomy than the general person)
Also, the AI has picked up the notion of a crown of thorns or a halo around the martyr's head but seems to often merge it with the hair.
Again it doesn't have a model of what a halo or thorn-crown is as an object and a symbol, the data just contains lots of examples with 'ring-like structure close to / around / above the top of the head'. So with a certain probability it tries something like that.
Things melting into each other is a bit of a general problem.
All of this of course can and will be fixed by filtering the generated images but even if these obvious machine-typical errors aren't apparent anymore it is still a work of 'blindsight'...
I also wonder how good these models can be about specifics. Like, if you asked for every woman to have her hair in a bun or ponytail, instead of down loose, would it manage? The truth is, no doubt, that this kind of development is only a few years away regardless.
I guess so long as there are enough images of the hairstyles categorized with the right names in the dataset it will work very well.
However someonoe could describe to you, without knowing its name, a certain foreign/historic style you've never drawn before, and you could work out roughly what the result should look like. The AI model just has data about different looks associated with the terms but can't 'build' one.
As for the impact on art as such ... I guess the impact will be very rapid for mass-produced imagery that no one really looks at too closely. So the random shiny happy people that are portrayed in the background of products/ads will soon be all AI-made. Same for random stock photos of sharply dressed businesspeople for website banners and so forth. All art-hosting websites will have to deal with the issue of getting flooded with such imagery.