Borrow the skeleton, not the picture.
A thumbnail that works usually works because of where things sit, what the eye reaches first, and how it is lit. Re-Prompter reads that, and hands it back as something you can build on with your own content.
What it reads
Structure, and deliberately nothing else.
The vision model is instructed to describe arrangement and light, never the specific people, brands, products or wording, and never to transcribe the text it can see. Colours are measured off the pixels rather than guessed, so the palette is right even when the model is vague.
Arrangement
Layout in a sentence, subject count, primary and secondary positions, camera angle, background.
Light and colour
Lighting description and the dominant colours, measured from the image itself.
Hierarchy
What the eye reaches first, second, third — and the devices doing it: rim light, directional gaze, an arrow, a scale jump.
Typography placement
Where the headline sits and how large it is relative to the frame. The wording itself is never taken.
What you control
Five switches and a slider.
Preserve layout, colour feel, lighting, subject position and typography placement, independently. Then the originality slider decides how much of any of it survives.
Close
Same skeleton, same reading order, entirely different content.
Balanced
The reference structure guides it, the content is yours.
Loose
Same energy and lighting, its own arrangement.
Mood only
Shares the feeling and little else.
Where the line is
This is not a copying tool, and it does not pretend to be one.
Re-Prompter will not reproduce a reference pixel for pixel, will not lift its wording, and will not carry across a recognisable person or brand. Every generated result carries a note saying it was built on the structure of a reference and is not the original. That constraint lives in the prompt and in the feature, not in a paragraph of terms nobody reads.