GPT Image 2.5 repeats the headline: say how many times it appears

When GPT Image 2.5 prints your headline twice, or adds a tagline you never wrote, the fix OpenAI documents is to quote the required copy and say how many times it should appear. Add an instruction for no extra text, then check the result. On Sume, send that prompt to openai/gpt-image-2.5 and use n to sample a few candidates, knowing each completed image is billed.

The instruction OpenAI gives

OpenAI's image prompting guide says to quote the required copy and specify how many times it should appear, so the text is not duplicated. The same page says to ask for no extra text and then check spelling and legibility.

Those are two separate constraints, and models tend to break them separately. One controls repetition. The other controls invented words.

A prompt with a count in it

Write the count as a plain sentence, next to the text itself, not buried at the end: Headline text, exactly once: "SPRING SALE". No other text anywhere in the image. Then describe position and typography in a separate sentence, so the count cannot be read as a layout hint.

Keep the copy short. OpenAI's image generation guide notes that the model can still struggle with precise text placement and clarity, so a three-word headline is a safer test than a paragraph.

Sample candidates with n, and count the bill

Sume's Image API accepts n to request several images in one call. The docs cap it at 10 per call, and say per-model ceilings are lower, so read the n range descriptor from GET /v1/images/models for the model you use.

Billing is all-or-nothing per image: a completed generation is billed in full, a failed or cancelled one is not. The docs say the endpoint pricing line is what is charged to your wallet, so cost_usd multiplied by n is the cost of a sample run. Four candidates cost four images, whether or not the headline is right on any of them.

Because every candidate is billed, test the wording at quality low first if the model's catalog row lists it, then run the final at the quality your text needs. A draft-then-final workflow covers that cost trade.

What this does not fix

A count instruction lowers the chance of a repeat; it does not remove it. If a layout needs exact text, exact placement and a fixed brand font, render the image without text and set the type in your own design step. That is a limit of image models in general, not something Sume changes: the API returns the image the model drew.