An image model only knows what it's told. "Woman in a blue dress" gets a generic picture; naming the fabric, the light direction, the camera height, the skin texture and what must stay unchanged gets something closer to a brief. Good prompts read like a photographer's shot list, decided down to the number of objects in frame.
It is also where most of the trial and error lives. Small wording changes swing the result, a model may ignore half the instructions, and what works for one product fails on the next. Teams that run AI themselves often spend more hours rewording than the shoot would have taken.
