AI model for product photos on Sume: Auto or a pinned id

Sume does not name a single best image model for product photos, and its docs publish no quality ranking to lean on. What it gives you is a choice of mode. Send model: "sume/auto" when you want a usable picture without choosing, and pin a catalog id when you need the same model every time, a mask, or several reference photos of the product.

The mode differences below come from the Image API docs and the catalog code, read 2026-09-29. They describe request behaviour, not how good the pictures look.

What does sume/auto give me for a product photo?

Sume picks the model family for you and never discloses which one ran. sume/auto is not in GET /v1/images/models, and job.model stays sume/auto in the result. That is convenient for one-off shots and awkward for a catalog of hundreds of SKUs, because you cannot say later which model made a given image or ask for it by name.

When should I pin a model instead?

Pin one when the shot depends on inputs Auto cannot promise you. Each row below is a documented capability, not a quality claim.

How do I keep the product identical across a set?

Pass the product photo as a reference and ask for the edit on an edit-capable id. On edit calls, use aspect_ratio: "auto" so the output follows the reference; leaving it out is not the same. A model that is text-only, such as the Imagen 4 ids, refuses reference images, so check the endpoint record first.

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": "Keep the bottle unchanged; place it on a marble counter",
    "input_references": ["https://example.com/bottle.png"],
    "aspect_ratio": "auto"
  }'

How do I compare candidates fairly?

Run the same reference and prompt against two or three pinned ids, keep the ratio fixed, and read the cost on each result, which is the USD amount billed to your wallet. Then judge the images against your own product, since the docs offer no ranking. If you settle on Auto for speed of setup, keep in mind that Sume does not disclose which model ran, so re-check a sample of results whenever you run a large batch.

What if a parameter is refused?

A model accepts only the values its catalog descriptors list, and anything else returns 400 unsupported_parameter. Read supported_parameters on the endpoint record before you pin. The wider model table is in which image model to use, and reference handling is in image generation with reference images.