This is pretty experimental, but can be a lot of fun if you persist past the duds. The idea was to create scenes that have elements that are obviously, yet skillfully photoshopped or inpainted in. Flux wants to fix everything during training, so it fought the concept pretty hard.

The trigger photoshopped was trained in, but you can also just describe two scenes. In a lot of my examples I've described a scene at the top of the image and a separate scene at the bottom. You can also use a phrase like morphing into. Broad items like umbrellas and hats are convenient transition areas between scenes.

Actually, one tremendously helpful tip I learned from watching @AIDigitalMediaAgency use this is to prompt for many different types of scenes in specific places. So, for example, a woman's face that has eyes morphing into a waterfall, erupting volcanoes photoshopped into her cheeks, hair shaped like snowy mountain peaks, etc. Even though I trained on images that only had two distinct scenes, the model can generate lots of them.

Limb disfigurement is fairly prominent. Sorry 'bout that.