I saw someone using JSON-structured prompts for image generation before, turns out it works just as well for image editing.
The logic is pretty simple: natural language is inherently ambiguous. You say “change this, edit that,” and the model might misinterpret. Switch to a clean JSON structure, list out the fields you want to change one by one, and the ambiguity gets crushed. With models like Flux.2, using structured prompts gives you decent results on the first try, and you end up re-rolling way fewer times.
Feels like overkill at first, but for scenarios where you need precise control, it’s a lifesaver—especially when you only want to tweak one thing and keep everything else the same. I use this trick all the time when editing design drafts, saves me a ton of hassle.