People were already using image-input capabilities of some previous AI art models to do iterative processes like this, but the examples I’ve seen in the past few days with the new Stable diffusion img2img are really impressive. Especially the ones that start with really crude sketches like these (There are others using photos, 3d renders, etc as original inputs)

I haven’t tried it yet myself but my impression is that the input images are combined with the usual text prompt stuff (ex. “portrait of woman, realistic detailed anime ArtStation”), then the output is fed back in iteratively.

Manual editing can be used to further refine in-between iterations, make composition adjustments, do color underpainting, collage and stitch together multiple images panorama style (such as a face produced by an AI art model specialized just for faces) etc etc

Some devs are working on directly including collaging/in-painting or uncrop/out-painting (like in the bottom example) capabilities in their tools as well

Some more at #img2img tag on twitter https://nitter.kavin.rocks/search?q=%23Img2Img