A general prompting pattern for models with in-model image editing (e.g. GPT Image 2): instead of writing a fresh prompt for a new variant, feed the prior generation back in as the reference image and explicitly state which elements must stay identical (face, outfit, pose) and which should change (background, lighting, angle). This shifts the model's task from 'generate something new that resembles X' to 'edit X, holding these parts fixed,' which produces tighter consistency than re-describing the whole scene from scratch each time.
This is distinct from pre-generation reference prep techniques like Clean Your Plate (Reference Image Prep for Video Generation) (cleaning a plate before it ever enters a video model) or asset-sheet techniques like Three-Panel Character Sheet and Prop Sheet for Cross-Style Object Consistency (building a fixed reference asset once, up front). Those front-load consistency into a prepared reference; this technique instead treats each successive output as the next input, iterating forward one edit at a time. It pairs naturally with Color Transfer / Palette Locking and Blocking Lock via Position Reference as another lever for constraining what an edit is allowed to touch.
Из тем: Unsorted, Holding Continuity Across Independently Generated Shots