AI image with text generator: get the words right

To generate an AI image with text, put the exact words in the prompt inside quotation marks, say where they go and how they should look, and use a model with a quality setting turned up. Then proofread every letter before you use the image. When the words must be exact, such as a price, a date, or a brand name, generate the picture with empty space and set the type yourself in a design tool.

Sume facts come from the Image API and Image 1.0 docs and the model catalog that GET /v1/images/models serves, read on 2026-09-27.

How do I write a prompt that puts text in an image?

Treat the words as part of the scene description, and be specific:

• Quote the exact words: a shop sign that reads "Open 7 days".
• Keep it short. A headline of a few words is easier to get right and to check than a paragraph.
• Say where the text sits, how big it is, and its color and lettering, such as bold white sans-serif across the top.
• Ask for several versions in one call with n, and keep the one whose lettering is correct.

Which quality setting helps with text?

Raise quality on a model that lists it. Sume's page for Image 1.0, a retiring alias of the Auto path, says to escalate quality "for finals, dense text, or packaging". On POST /v1/images, four catalog models list the field today; sending quality to any other model returns 400 unsupported_parameter.

Higher settings cost more on ChatGPT Image 2.5, whose price is estimated from size and quality. High quality is also one of the slow configurations most likely to return 202 with a job to poll instead of the finished image.

How do I add or change text in an existing image?

Send the image as a reference to a model that edits from references, such as ChatGPT Image 2.5, and write the new words in quotes, with their position. Add aspect_ratio: "auto", which the docs recommend on edits to match the reference's shape, and host the image at a public HTTPS URL.

The model redraws the whole image, not only the letters, so compare the result with the original: faces, logos, and other text can change too.

curl -X POST "https://api.sume.com/v1/images" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "Replace the words on the shop sign with \"Open 7 days\" in the same lettering style. Keep everything else the same.",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/storefront.jpg" } }
    ],
    "aspect_ratio": "auto",
    "quality": "high"
  }'

When should I set the text myself?

Whenever a wrong letter would cost you: prices, dates, legal lines, brand and product names, and long copy. Ask the model for the picture with room for the words, such as "leave the top third as clear sky for a headline", then add the text in a design tool where you can check every character.

AI magazine cover generator uses the same split for mastheads, and add text over a video covers words on moving images.

What are the limits of AI text in images?

• The words are part of the pixels. You get a flat png, jpeg, or webp file with no editable text layer, so a typo means a new generation or a fix in an editor.
• No setting guarantees spelling. Check every word, including any extra lettering the model adds on its own.
• Results come back as Sume-hosted, signed URLs in data[].url, so save the final version in your own storage.