AI photo editing prompts: change one thing, keep the rest

A good AI photo editing prompt has two parts: what to change, stated concretely, and what to keep, listed. For example: “Change the sky to a clear sunset; keep the people, their faces, the buildings, and the framing the same.” Send the prompt with your photo to an image model that edits from reference images, change one thing per request, and start each new edit from your original photo rather than from an edited copy.

On Sume, an edit is one POST /v1/images request with your photo as a reference. The facts below come from the Image API docs, the retiring Image 1.0 page, and the request schema in the Sume API reference, read on 2026-09-28.

How do I write an AI photo editing prompt?

Use the same two-part shape for every edit:

• Change: one thing, in concrete words. Not “make it better”, but “warm golden-hour light from the left”.
• Keep: everything that must survive, by name: faces, text, logos, products, the pose, the framing.
• Leave out a description of the rest of the photo. The model already has the photo.
• With two photos, say which is which by what each one shows. A reference carries only its type and URL, with no role field.
• Sume's docs show two edit prompts: “make this scene look like a watercolor painting”, and, on the retiring Image 1.0 page, “Keep the product identical; swap the background to a soft daylight studio.”

What are some AI photo editing prompt examples?

One per kind of edit. They are illustrations to adapt, not wording any model is documented to follow:

• Background: “Keep the person and their clothes the same; replace the background with a quiet beach at sunset.” More in how to change the background of a photo with AI.
• Lighting: “Relight the photo with soft, warm window light from the left; keep the people, the room, and the framing the same.”
• Color: “Change only the car to matte black; keep the wheels, the street, and the reflections the same.” More in how to change the color of an object in an image.
• Style: “make this scene look like a watercolor painting”, the docs' own example. More in photo to painting AI.
• Season or time of day: “Make it a snowy winter afternoon; keep the house, the trees, and the framing the same.”
• Outfit: “Change the man's T-shirt to a white linen shirt; keep his face, his pose, and the background the same.”
• Cleanup: “Remove the trash can on the left and continue the brick wall behind it; keep everything else the same.” The docs don't describe object removal, so treat this one as a try and check the patched area.

How do I send the edit to an image model?

• Put the photo at a public HTTPS URL and send it in input_references to a model that edits from references, such as openai/gpt-image-2.5. A text-to-image-only model rejects it.
• Set aspect_ratio: "auto" to keep the photo's shape. Leaving it out is not the same as auto.
• Set n for several takes of the same edit, up to 4 on ChatGPT Image 2.5 today, and keep the closest match.
• The call waits up to 30 seconds and returns 200 with the images; a slower one returns 202 with a job to poll.

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": "Change the sky to a clear sunset. Keep the people, their faces, the buildings, and the framing the same.",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/street.jpg" } }
    ],
    "aspect_ratio": "auto",
    "n": 2
  }'

Why start every edit from the original photo?

Because each edit is a new drawing of the whole photo. Faces, small text, and fine texture can shift a little in every pass, so an edit of an edit can drift further from what you shot. Find the wording for each change one request at a time, from the original. When you want two changes in one picture, put both working wordings into one prompt and send it with the original.

Can I edit just one area, and what are the limits?

Only on ChatGPT Image 2.5, and only partly documented: it takes an optional mask_url on edits, but the docs don't say what the mask image should contain, so name the area in the prompt instead. The other limits: