I wonder how tweakable those generated images are. The most important requirement for 'game art creation' is that it is created in a tight feedback loop between an artists, an art director who needs to enforce an overall artistic vision, and the cold, hard requirements of the game design.
Can I tell the AI "that looks great, keep the trees and house, but I need the door on the right side, and the sky a bit less cloudy". Will it be able to incorporate such instructions without "remixing" a very different image? Can it "understand" the suggestions the same way a human can without having to go into so much details that it essentially becomes procedural generation? Is this even something that's possible with the current approach?
If that's not possible, then I can just as well google and hope to find a matching image.
I’ve been on the beta testing for Stable Diffusion and one thing that they have is give you the seed of the image with which the space is populated. What this enables is enabling you to tweak the image, as you can keep the ‚base‘ image and iterate on the prompt.
It’s not as sophisticated as you described, yet, but close.
In addition I can imagine that it is a matter of the training set. As of now, the AIs seem like jack-of-all-trades, so they know a bit of a lot of different styles and topics. But with Stable Diffusion being able to run locally, you could specialize it by training it with high detail on, let’s say, landscape photography. So then one might be able to direct the AI more precisely.
Can I tell the AI "that looks great, keep the trees and house, but I need the door on the right side, and the sky a bit less cloudy". Will it be able to incorporate such instructions without "remixing" a very different image? Can it "understand" the suggestions the same way a human can without having to go into so much details that it essentially becomes procedural generation? Is this even something that's possible with the current approach?
If that's not possible, then I can just as well google and hope to find a matching image.