Image to image AI: turn your photo into a new picture

Image-to-image AI makes a new picture from a picture you give it plus a text instruction. Instead of starting from words alone, the model takes your image as a reference and redraws it: restyled, with a new background, recolored, or as a variation of the same subject. With Sume's image API, it is the same POST /v1/images call as text-to-image, with your photo's URL in input_references, sent to a model that accepts image input.

Sume facts come from the Image API docs and the model catalog that GET /v1/images/models serves, read on 2026-09-28. The full request, field by field, is in image generation with reference images.

What can image-to-image AI do with my photo?

The prompt says what to change, and the reference supplies what the model should work from. Common jobs:

• Restyle: turn a photo into a painting or an illustration. The docs' own example prompt is "make this scene look like a watercolor painting". More in photo to painting AI.
• Replace the background. The docs' product example reads "Keep the product identical; swap the background to a soft daylight studio". See how to change the background of a photo.
• Recolor one object, such as a jacket, a car or a sofa, as in change the color of an object.
• Make variations: the same subject in a new pose, setting or season.
• Carry one character or product into new scenes by sending the same reference every time, as in consistent character AI image generator.

How is image-to-image different from text-to-image?

Text-to-image starts from the prompt alone. Image-to-image adds one or more images the model must work from, so the result starts from your picture's subject and layout instead of an invented one.

On Sume, a text-to-image-only model refuses a photo with 400 unsupported_parameter, and image generation API models shows how the catalog marks each model. The models below take your photo in input_references, and in current code each one sends it to a dedicated edit mode:

How do I run image-to-image with an API?

Put the photo's public HTTPS URL in input_references, say what to change in prompt, and pick a model from the table. The request below is the docs' example with aspect_ratio: "auto" added, which the docs recommend on image-to-image calls so the output matches the reference.

• Localhost, private-network and non-HTTPS URLs are rejected before submission, so host the photo at a public HTTPS address.
• Leaving aspect_ratio out is not the same as sending "auto". The Seedream models don't list auto, so give them a ratio from their list.
• The call waits up to 30 seconds. A 202 instead of 200 means the job isn't finished: poll GET /v1/jobs/{id}/status until it is completed, failed or canceled, and read GET /v1/jobs/{id}/result only if it completed.

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",
    "prompt": "make this scene look like a watercolor painting",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/photo.jpg" } }
    ],
    "aspect_ratio": "auto"
  }'

Will the new image keep my photo's details?

Not exactly. The output is a new image, not your file with one area patched, so faces, small text, logos and fine patterns can drift even when the prompt asks to keep them. Name what must stay the same, change one thing per request, and compare each result with the original. Prompt patterns for this are in AI photo editing prompts.

What does it cost, and what are the limits?

• Each model has a per-image rate in the pricing line of GET /v1/images/models/{model_id}/endpoints, plus a 5.5% agent fee by default. On ChatGPT Image 2.5, that price is estimated from tokens, so the output's size and quality and the reference images you send all move it. usage.cost in the response is the USD amount billed, and a failed or cancelled generation is not billed. AI image generator API cost lists the rates.
• Seedream 4.5 returns one image per call when you send references, whatever n says (current code).
• Result URLs are Sume-hosted and signed, so download the images you keep.