How to generate variations of an image with AI

To generate variations of an image, give an AI image model the original as a reference, say in the prompt what may change and what must stay, and ask for several results at once. Each result is a new generation, so every run gives a fresh set; to explore further, feed your favorite back in as the next reference.

On Sume, that is POST /v1/images with the original in input_references and n set to the number of variations. The details below come from the Image API docs and the catalog that GET /v1/images/models serves, read on 2026-09-27.

How do I get several variations in one call?

Send the original as a reference and set n. The docs allow up to 10 images per request and say per-model ceilings are lower; on ChatGPT Image 2.5, ChatGPT Image 2, and the two Nano Banana models, the ceiling is 4 today.

• aspect_ratio: "auto" keeps the original's shape on models that list it, and leaving the field out is not the same as auto.
• Today an n outside the model's range is refused with 400 invalid_request.
• Four variations are billed as four completed images. A failed generation is not billed.

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": "A variation of the referenced poster: keep the layout, the bicycle, and the flat illustration style; change the color palette and make the season autumn.",
    "aspect_ratio": "auto",
    "n": 4,
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/poster.png" } }
    ]
  }'

What should the prompt say?

The prompt is your control over how far a variation strays, so state both halves:

• Keep: the subject, layout, style, palette, or wording that must not change.
• Change: one or two things per run, such as colors, background, props, season, camera angle, or mood.
• Close copies: "Recreate this image with small changes to the background only."
• Wider exploration: name a new style or setting outright, and keep only the subject.

How many variations can each model return?

Seedream 4.5 is the exception today: its catalog lists n up to 4, but with a reference it returns one image per call, so send one call per variation. For the models that edit from a reference:

Can I control how different the variations are?

Only through the prompt. The request has no strength, similarity, or variation field, and in current code the body schema is closed: an unknown field such as strength fails the request with 400 invalid_request instead of being ignored.

seed, described as "Seed for deterministic generation", is in the schema but not served in v1, so sending it returns 400 unsupported_parameter. Keep the files you like, because a rerun makes a fresh set. More on seeds: AI image seed number.

What are the limits?

• The original must be at a public HTTPS URL, and a text-to-image-only model rejects references.
• Slow configurations, a large n among them, are the most likely to answer 202 with a job instead of 200 with the images. Poll GET /v1/jobs/{id}/status, then fetch GET /v1/jobs/{id}/result.
• Result URLs are Sume-hosted and signed. Download the variations you keep.
• For several images from a text prompt alone, see 4K AI image generation: resolution, quality, and image count; for reference mechanics, Image generation API with reference images.