Nano Banana 2.1 mask editing: Sume's mask_url is GPT-only

Nano Banana 2.1 was announced on 2026-10-06 with sharper mask-based editing: you mark one region and only that area changes. On Sume, the mask_url request field is accepted only by the two ChatGPT Image 2.5 variants. If you send mask_url with google/nano-banana-2.1, you should expect a 400 unsupported_parameter, not a silent drop.

What Google describes

The launch coverage describes the mask feature at the product level: mark a region, change only that region, and keep the rest. It does not give an API parameter name in the part I read, so this post makes no claim about how Google's own API takes a mask.

What Sume accepts today

The Sume Image API docs are explicit. mask_url is a public HTTPS mask URL, and it is live only on openai/gpt-image-2.5 (Flare) and openai/gpt-image-2.5-sunburst. The request table says the catalog controls validation: if a model does not list a parameter in supported_parameters, the call is rejected with 400 unsupported_parameter.

What to do for a Nano Banana edit on Sume

Describe the region in words and send the original as an input_references item. Add a preserve clause to the prompt, such as keeping everything outside the named region unchanged. It is a text instruction, not a pixel guarantee, so check the result against the original before shipping.

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-2.1","prompt":"Change only the sofa colour to deep green. Keep the room, lighting and every other object unchanged.","aspect_ratio":"auto","input_references":[{"type":"image_url","image_url":{"url":"https://media.sume.com/uploads/room.jpg"}}]}'

When a real mask is the right tool

Use openai/gpt-image-2.5 with mask_url when the edit region must be exact, for example a single label on a pack shot. On edit calls, aspect_ratio: "auto" matches the reference; omitting the field is not the same as auto.

The reference URL must be public HTTPS. Sume rejects localhost, private-network and non-HTTPS URLs before submission.

Decision rule

A short rule for choosing the model for an edit:

• Exact region, pixel-level control: openai/gpt-image-2.5 with mask_url.
• Whole-scene change with a preserve clause: Nano Banana 2.1 with a reference.
• Not sure: run both once and compare the changed pixels against the original.