JSON Aesthetic Prompting

JSON Aesthetic Prompting is a technique I’ve been exploring for getting more precise, repeatable control over AI image generation. The basic idea: instead of writing long, vague text prompts and hoping the AI interprets your intent, you define your aesthetic as structured JSON data — think material properties, lighting conditions, surface textures, color palettes — and feed that structured definition into an image generator alongside your source asset.

How it works

I started with a simple ServiceNow star icon and wrote a JSON object describing a target aesthetic: hyperrealistic rendering, frosted crystal material, specific lighting angles, the whole deal. When you feed that JSON definition plus the source icon into an AI image generator, it retextures the icon to match that aesthetic with surprising accuracy. The cool part is it works in reverse too — you can take a generated icon and ask the AI to flatten it back to a clean 2D version. You can even hand it an existing image you like and ask it to output the JSON aesthetic definition for that image, essentially reverse-engineering a style into a reusable data object.

Why this matters

The real use case here is scale. If you’re working on an icon system and a stakeholder asks to see ten different style directions across twenty icons, that’s two hundred assets to mock up before anyone’s made a decision. With this approach, you define each direction as a JSON aesthetic, run your icon set through each one, and have a massive spread of explorations ready for review in a fraction of the time. You’re not replacing the craft of manual refinement — you’re front-loading the exploration phase so that when you do invest hands-on time, you already know which direction has legs. It turns what used to be a multi-day exploration sprint into something you can knock out in a single sitting.