The 7 Stages, Explained
What each stage controls, why it's a separate step in the pipeline, and the real photography, rendering, and design principles behind its controls.
What it controls
The structure picker (poster, product, portrait, logo, banner, illustration, custom), aspect ratio, visual style, background environment — and the canvas itself, where every element is placed, sized, stacked and written into by hand.
Why it's its own stage
Every later stage (typography, lighting, camera, texture) attaches to a specific element. Those elements have to exist, and have to be somewhere, before anything can be said about how they look. Choosing a structure seeds the canvas with that layout so the common cases start from a real composition rather than an empty rectangle.
Why the AI image model needs this explicitly
An image model generates the whole canvas in one holistic pass — it has no separate "layout" step of its own, and no memory of a page grid the way a design tool does. Left to its own defaults it will center a single subject and improvise where everything else goes. Because every element here carries real coordinates, the compiled prompt can state exactly where each piece of content sits as a percentage of the canvas, which gives the model's attention something concrete to bind content to instead of leaving composition to whatever is statistically most common in its training data.
Evidence & established principles
Modular grid systems (Swiss/International Typographic Style, Müller-Brockmann)
The safe-margin guides and snapping enforce a repeatable structure while you drag — the same discipline print design has used since the 1950s to keep dense information scannable instead of scattered.
Aspect ratio drives platform legibility
1:1 (square) suits feed posts, 9:16 (tall) suits Stories/Reels, 16:9 (wide) suits web headers — the canvas reshapes to the ratio you pick, so you compose for the real frame rather than discovering the crop later.
Direct manipulation beats description
Placing a title by dragging it takes one gesture; describing the same placement in prose takes a sentence and is still ambiguous. Only the words that actually appear in the image need typing.
Fields: imageType, aspectRatio, visualStyle, bgEnvironment, canvasMarginX/Y + the element list