Ask an image model for “a poster” and you get a poster-shaped image: picture on top, text at the bottom, both fighting for the same attention, nothing holding them together. It reads as generated because the composition is generic, not because the rendering is bad. The fix isn’t a better model. It’s giving the model a layout system instead of a subject. Below are two systems that reliably produce work that looks art-directed. Both come with fill-in templates. Both work in any current image model that accepts long prompts — Nano Banana, Midjourney, Seedream, GPT Image, Flux. Why layout instructions beat style adjectives Most poster prompts are a pile of adjectives: minimalist, elegant, high-end, professional, award-winning . Those words push the model toward the average of everything tagged that way in training. Average is exactly what you’re trying to escape. Layout instructions are different. “The bottom information zone occupies 25–40% of the height” is a constraint the model ...