It all heavily depends on what particular AI you are using and how you are using it. Currently the commercial AI engines are typically a bit limitted in how much you can adjust the result.
So one thing you can do is start out with a image, then ask the AI to mutate that image. This is most likey a bit too strong of an effect.
The second thing that you can do is ask the ai to describe the image. This however is a bit of a hit and miss. Generally prompting is the least able to do layout decissions one of the functions.
Next you search for a model that is specifically trained or modifed for the effect you are looking for. Something like this lora for stable diffusion should make it considerably easier.
Then you use something like control net, or style transfer for this. Both of these are more like the image transform but allow you to reconfigure the image more, such ss ignoring color or allowing you to use a sketch as a basis.
then you should consider just painting roughly over generated result and let the ai redraw the thing with moderate denoising.
If this fails you can train your own model. Three images is entirely plausible if you just modify an underlying model. Shouldnt take too much time.