GPT Image 2.5 text in the wrong place: fix it with a mask edit

When GPT Image 2.5 puts a headline in the wrong spot, do not regenerate the whole picture. Make a mask that covers only the text area and run one edit with mask_url; the rest of the image has the best chance of staying as it was.

OpenAI's image generation guide is direct about the limit: although text is "significantly improved, the model can still struggle with precise text placement and clarity." That matches what the Sume Image API offers for fixing it: an optional mask_url on ChatGPT Image 2.5 edits.

What does Sume accept for this edit?

The Image API docs say openai/gpt-image-2.5 (Flare) and openai/gpt-image-2.5-sunburst both support up to 16 image references, an optional mask_url, and background: auto|transparent|opaque. The mask URL must be public HTTPS, like every reference.

OpenAI's guide says masks identify the areas to replace, and that the image and mask must be the same format and size, under 50MB, with an alpha channel. Build your mask to that rule before uploading it somewhere public.

How do you write the prompt?

Name the text exactly, say where it goes, and say what must not change. Keep the quoted copy short, because longer strings are where placement and spelling slip first.

• Quote the string: Headline reads "SUMMER SALE" in straight quotes.
• Describe the zone: top third, centred, white sans-serif on the dark sky.
• Add the preserve line: Keep the product, background and lighting unchanged.
• State the count: say the text appears once.

A request you can send

Replace the two URLs with your own hosted image and mask. The model id comes straight from the docs.

Two practical notes on the call. First, keep aspect_ratio at auto so the edit keeps the original's shape; the Image API docs say omitting the field is not the same as auto. Second, pick the quality deliberately. Omitted quality defaults to high on Sume, and auto reserves the max amount, so a text-only touch-up does not need either. Try medium first and move up only if the letters are soft.

import os, requests
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
body = {
    "model": "openai/gpt-image-2.5",
    "prompt": 'Headline reads "SUMMER SALE" once, centred in the top third, '
              "white sans-serif. Keep the product, background and light unchanged.",
    "input_references": [{"type": "image_url",
        "image_url": {"url": "https://example.com/ad.png"}}],
    "mask_url": "https://example.com/ad-text-zone-mask.png",
    "aspect_ratio": "auto",
    "quality": "high",
}
r = requests.post("https://api.sume.com/v1/images", headers=H, json=body, timeout=60)
print(r.status_code, r.json())

What should you expect from the result?

A mask is guidance, not a hard clip. Our earlier post on edits that leaked outside the mask explains why, and the pixel-diff post shows how to measure it. Plan on checking the result once rather than trusting it.

If the text is still wrong after two attempts, stop paying for retries. Drop the text from the image, render it in your design tool, and use the model for the picture only. That is the honest answer when exact typography matters, for example a legal line or a price.

A final tip on building the mask. Make the white (edit) zone slightly larger than the old text, not tight around it. A tight mask leaves ghost remnants of the previous letters at the edges, while a generous one gives the model room to repaint the background behind the new words. Remember the OpenAI rules from the checklist post: same size and format as the image, with an alpha channel.