Image-to-image strength or denoise: no field on Sume, do this

Sume's Image API has no image-to-image strength or denoise field. A request that sets a parameter the selected model does not list is rejected with 400 unsupported_parameter instead of being silently dropped, and seed and output_compression are in the schema but advertised by no model (Image API). To control how far the output moves from your reference, control it in the prompt, the model choice, and the mask.

Four levers that do exist

Reference-guided models on Sume follow instructions about what to keep. That makes the prompt your strength dial: the more specific the preserve list, the less the image changes.

• Preserve list: name the things that must stay, such as pose, framing, logo, face and colours.
• Change list: one sentence on what changes. Edits that ask for one change drift less than edits that ask for five.
• Mask: GPT Image 2.5 takes mask_url, and the mask is guidance, not an exact shape, so check edges (OpenAI notes this).
• Model: pick one that lists input_references for your use, and read its descriptors before you code against it.

Check which fields a model takes

Do not hard-code a field list. Read supported_parameters from GET /v1/images/models/{model_id}/endpoints and send only what is there. If your old pipeline sent strength, drop it in the Sume path and add a preserve list instead.

curl "https://api.sume.com/v1/images/models/openai/gpt-image-2.5/endpoints" \
  -H "Authorization: Bearer $SUME_API_KEY"

Control and repeatability

Without a seed, two identical requests give different images. If you need a stable result, keep the one you like and use it as the reference for further edits instead of re-rolling. Sume bills each completed generation, so a re-roll is a new charge.