Nano Banana 2.1 tracks 4 characters, 10 objects: Sume's cap

Google reports that Nano Banana 2.1 accepts up to 14 reference images and keeps track of up to 4 characters and 10 objects. Sume's Image API caps input_references at 10 per request (16 on ChatGPT Image 2.5), and each model's own ceiling is in its catalog row. So a 14-reference Google workflow needs to be cut down before it runs on Sume.

The numbers side by side

The Google figures come from launch coverage published on 2026-10-06, and the Sume figures come from the Image API docs. The per-model ceiling on Sume is the input_references range descriptor, which you read before you build the request.

How to fit 14 references into 10

Group by subject rather than by shot. Four characters need one clean reference each, which leaves six slots for objects, a style frame and a background. Merge small props into a single flat-lay image that shows several objects at once. A merged image costs one slot instead of three.

Name each reference in the prompt by order: image 1 is the lead character, image 2 the product, and so on. The docs do not describe reference-order semantics for Nano Banana, so test the ordering on two or three prompts before you run a batch.

Check the descriptor, then send

Read the catalog row, then post the request. A request with more references than the model allows is rejected rather than trimmed.

curl "https://api.sume.com/v1/images/models" \
  -H "Authorization: Bearer $SUME_API_KEY"

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":"Image 1 and image 2 sit at a cafe table, image 3 is the mug on the table. Keep faces and the mug logo identical.","input_references":[{"type":"image_url","image_url":{"url":"https://media.sume.com/uploads/a.jpg"}},{"type":"image_url","image_url":{"url":"https://media.sume.com/uploads/b.jpg"}},{"type":"image_url","image_url":{"url":"https://media.sume.com/uploads/mug.jpg"}}]}'

Honest limits

The Sume docs do not claim identity tracking numbers for any model, so the 4 and 10 are Google's claims about Google's endpoint. Sume also does not run Google Search grounding for this model according to the docs; no grounding parameter is listed. Verify consistency on your own subjects with a handful of runs, since Google's Elo comparison was not independently verified at launch.

A packing checklist

Before you send a crowded scene:

• Give each character one clean, front-facing reference.
• Combine small props into one flat-lay reference image.
• Keep the total at or below the input_references max in the catalog row.
• Describe each reference by its position in the prompt.