I saw talk the other day how these ML art models aren't really suited to doing something like illustrating a picture book because it can synthesize a character once but wouldn't be able to reliably recreate that character in other situations.
Didn't take long for someone to resolve that issue!
This doesn't feel too far off. With the img2img stuff you can give a picture of a character and the tool can spit out new images of that character or transpose them in to new art styles.
It doesn't feel like this tech has hit any hard limits on things that are impossible or very far away yet. Every limitation seems to be getting broken at rapid pace.
It's not quite at that level yet. The paper introducing it recommends using only 5 images as the fine-tuning set so the results are not yet very accurate.
Didn't take long for someone to resolve that issue!