Translate text inside an image with GPT Image 2.5 edits

To translate the text in an image with GPT Image 2.5, send the image as a reference, give the new wording in quotes, and add the line OpenAI's cookbook uses: do not change any other aspect of the image. On Sume that is POST /v1/images with openai/gpt-image-2.5, one reference, and aspect_ratio set to auto so the frame is not reshaped.

What OpenAI documents for this edit

The OpenAI Cookbook prompting guide lists translating text as an edit case and says to add that nothing else in the image should change. The image prompting page says to quote required wording, list what must stay the same (identity, geometry, layout, lighting, labels) and ask for no extra text.

The Sume call

A reference must be a public HTTPS URL. For edits, Sume's docs advise aspect_ratio: "auto" to match the reference, and note that omitting the field is not the same as auto.

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",
    "aspect_ratio": "auto",
    "quality": "high",
    "prompt": "Replace the sign text with \"OUVERT\". Do not change any other aspect of the image. No extra text.",
    "input_references": [
      {"type": "image_url",
       "image_url": {"url": "https://example.com/storefront.png"}}
    ]
  }'

What can go wrong

The model redraws the whole image, so even a correct translation can shift a logo or a face. OpenAI's image guide says the model can struggle with text clarity and with keeping recurring brand elements consistent, which is the combination a translated sign or ad hits.

The OpenAI pages we read do not list which writing systems render reliably, so test each target language on a sample before a batch. Check accents, spacing and line breaks on every output.

Limit the area with a mask, if you can

Sume accepts an optional mask_url for GPT Image 2.5 edits. OpenAI says masking with GPT Image is prompt-based: the mask is guidance and the model may not follow its exact shape. Use it to narrow the edit, and still keep the unchanged-pixels instruction in the prompt. See how a mask edit works on Sume.

For a batch of translated variants, remember that each completed image is billed in full and failed ones are not, so run one language end to end before fanning out.