Reverse-engineering photo style into a JSON prompt for realistic UGC images
The workflow uses GPT-6 Astra to turn a reference photo into a JSON prompt describing lighting and color grading, then feeds it to GPT Image 2.5 to generate product shots with a fixed character for consistency.
how to generate ai photos that look 100% real with gpt-image 2.5 2.5 is a step up from 2 but it carries the same tells straight out of the box with no reference you still get that ai look so stop describing the aesthetic and start stealing it find a photo that already has the look you want drop it into gpt-6 astra and ask this "analyze this photo and give me a very detailed json prompt describing its visual style. break down the color grading, the lighting and the exact colors so i can apply that look to my own images" astra reverse engineers the reference into a prompt you were never going to describe that grade yourself take the json into gpt-image 2.5 attach your product and prompt "using this json as reference, generate a person holding my product" then generate the same thing without the json and put them side by side the lighting and the grade are the whole difference save the character you like attach that photo on every future generation and ask it to preserve the same facial features now you have a consistent person across your entire product library want a different mood go back to astra, iterate the json, run it again the tells are almost always the same two things flat default color and that plastic grain so grade on purpose and check those first the json controls the light the character controls consistency you control the product that is a repeatable ugc pipeline instead of rolling the dice every time