If any if the training set is public domain or creative commons, I'd argue that this make the image at best an integration work, not derivative nor fair use, and as bits from the images are indistinguishable from each other, you shouldn't be able to license it.

Like code generated from models trained on AGPL/GPLv3 licensed code should also be AGPL/GPLv3.