Nano Banana 2 reference slots (10+4+3) vs Sume's flat reference list

Google's Nano Banana 2 splits references into slots: up to 10 object, 4 character and 3 style images. Sume's input_references is one flat array with a maximum of 10 on this model family, and Sume does not map Google's slots. If you rely on the character or style slots, you cannot reproduce that split through Sume.

Reference limits

Limits come from Google's image-generation page; the Sume limit comes from the Image API docs.

The arithmetic

Adding the NB2 slots gives 10 + 4 + 3 = 17 images, more than the 10 that Sume accepts in one list. Even if you put every slot's image in the flat array, you would have to drop 7 of them, and the model would get no label telling it which are characters or styles. The Sume docs describe no such label.

What to do on Sume

On Sume, the array is checked against the model's catalog descriptor. Read input_references in supported_parameters before you send a large array. Reference URLs must be public HTTPS.

• Use the prompt to say which image is a character and which is a style.
• Keep the array to the images that matter most; prefer fewer, cleaner references.
• If you need the 10+4+3 split, call the Google API directly.
• Quote the billed price from pricing: google/nano-banana-2 is $0.08 list x 1.25 = $0.10 in the repo's static table.