AI generated art

[KYWizard]

Out of curiousity I have used some of these to try and generate fantasy maps, and it is pretty terrible.

Anyone have any thoughts on this?

Full disclosure I am a new member and just looking to make some posts on here. Figured this might be an interesting topic.

[Redrobes]

We had some discussions and maps made using the new tools here:
https://www.cartographersguild.com/s...ad.php?t=51132

As it stands they are pretty terrible at making them. Either they rip off existing works to a degree which is too close to be novel or they can transform style to some shapes in a sort of fantasy sense.

But a map has qualities about it that are consistent across the page and make sense which no AI generator has managed without coding in some rules about what its trying to make. And secondly, fantasy covers such a diverse set of styles its not like asking for van gogh or impressionist or similar. Fantasy can mean almost anything. So its gonna be hard for AI to pin down what makes a fantasy map for a while yet.

I think some day its going to happen but I think fantasy maps will be at the latter end of the process and were quite some way off at present. I think they are a tool and can be brought to bear on making a map but to let it make the whole thing is out of scope for the time being.

[Azélor]

This Ai can't manage that yet.
Maps are complex, big and require consistency.
The Ai struggle a lot especially when you combine all of this.

One of the reason I suspect is that the AI was not trained to do maps.
It's is a niche art. The database is general and doesn't contain that many maps.
In order to get better results, you would need to train it with enough large quality images.
That is very ressource intensive (gpu) and not many people are willing or able to do it.

Small maps could be doable with enough tinkering.

[Falconius]

Maps are too inconsistent except for certain things like borders, or icons et cetera so machine learning can rip off those things somewhat consistently. The thing with maps is that they are usually defining a certain thing with visual information meaning that you'd have to go through and kind of teach the machine what all sorts of things were. Like they had to do for fingers or eyes or ears in regards to pictures. Maps don't work if you just mash a bunch of things that look right together, where as it does with art/pictures, because the brain does most of the work tying things together when it sees a person, and when you look closely the pictures seem to fall apart. With a map the brain isn't naturally trying to make it fit a certain understanding. I think if there was an machine learning project focused specifically on making maps it would do much better, but there is zero reason for anyone to make such a project. If you want machine generated maps procedural is the way to go.

[eViLe_eAgLe]

To add onto this, it is both a niche art, and to get a properly functional map you need data as well, things that the AI just doesn't have information to.

One of my favorite quotes that I don't recall the origin, is that maps are a mixture of art and data, with extremes on both ends.

If you want a map that you can add information to? Then... That's doable. If a bit crap in terms of consistency.
If you want a map wholesale? It's nigh impossible, I've tried with Midjourney and Stable Diffusion. I don't think it'll come close to that until we get properly functional multi-media AI's who generate things for us.

One of the more pertinent aspects I can imagine being used (and I have tried to use, but to little avail as it's always ever so slightly *off*) is asset generation.