Recreate an AI image with another model: what Sume Images carries over

To rerun an AI image on a different model in Sume Agents Images, use Recreate on the finished image. It restores the prompt, aspect ratio, quality, resolution, image count and up to 10 reference images from the original job, then clamps each setting to what the newly picked model supports. You can send the same request through the API.

What Recreate restores

The Images page builds a recreate state from the saved asset. Each field is normalized before it is shown in the composer, so the form never starts in a state the model would reject.

• Prompt: the saved prompt, or the title if no prompt was stored.
• Model: the model on the asset, or the default Auto if it is missing or legacy.
• Aspect ratio: the saved ratio if it is one of the known values.
• Quality: one of auto, low, medium, high, xhigh or max, clamped for the model.
• Resolution: 0.5K, 1K, 2K, 4K, 720p or 1080p where the model has tiers.
• Count: the number of images, clamped to the model's ceiling.
• References: only public media.sume.com HTTPS URLs without a query string, deduplicated and capped at 10.

Why settings change when you switch models

Models list different values. GPT Image 2.5 has six quality levels and Ideogram 4.5 has three. Grok Image allows one image per call and the others allow four. Recreate clamps rather than fails, so a request that used xhigh on GPT Image 2.5 will arrive on Ideogram 4.5 at a level that model lists.

Do the same through the API

Copy the fields and change model. If you keep the references, they must still be public HTTPS URLs. A media.sume.com URL from an earlier result works because Sume hosts it.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"google/nano-banana-pro","prompt":"Same prompt as the first run","aspect_ratio":"4:5","input_references":[{"type":"image_url","image_url":{"url":"https://media.sume.com/img/EXAMPLE/0.png"}}]}'

Compare fairly

For a fair comparison, change one thing: the model. Keep the prompt, the ratio and the references identical, and ask for the same n. Judge the outputs side by side, and compare the usage.cost on each response. If a model rejects a field, you get 400 unsupported_parameter instead of a silent drop, which tells you the setting did not carry over.

Before you recreate

Check that the reference URLs still open and that the new model lists references at all. Imagen 4, Recraft V4 and Qwen Image Max are text-to-image only in the catalog, so they take no input_references. See the route guide for where each call goes.