Nano Banana 2 takes 14 references; Sume's catalog allows 10

Google's Gemini API image guide describes Nano Banana 2 (model id gemini-3.1-flash-image) as accepting up to 14 reference images in total: up to 10 high-fidelity object images and up to 4 character-consistency images. That is the vendor's number. It is not the number you get through every router.

On Sume, google/nano-banana-2 is served with the shared reference ceiling of 10. The code that builds the catalog sets overrides only for the two ChatGPT Image 2.5 ids (16) and Ideogram 4.5 (5), so Nano Banana 2 and Nano Banana Pro stay at 10.

What happens at 11 references

Sume validates the count against the descriptor before it submits. The error is a 400 invalid_request with the message "accepts at most 10 input_references" and a max field of 10. Nothing is billed for a request that fails validation.

Three ways to stay inside 10

Rank your references and send the ten that carry the most identity. Characters usually matter more than props.

• Cut the least important references; keep faces and the hero product.
• Pack small props into one contact-sheet image so four items cost one slot.
• Switch to openai/gpt-image-2.5 when you truly need 11 to 16 references.

import os, requests
headers = {"Authorization": "Bearer " + os.environ["SUME_API_KEY"]}
refs = [f"https://cdn.example.com/ref-{i}.jpg" for i in range(14)]
body = {
    "model": "google/nano-banana-2",
    "prompt": "The same two characters at a picnic, product on the blanket",
    "input_references": [
        {"type": "image_url", "image_url": {"url": u}} for u in refs[:10]
    ],
}
r = requests.post("https://api.sume.com/v1/images", headers=headers, json=body, timeout=60)
print(r.status_code)

Check before you design around a vendor number

Reference counts, aspect ratios and tiers drift by provider. Read supported_parameters from the Sume catalog, not from the vendor page, when the number sets how many images you collect per shot.

How this was checked

Vendor facts come from the pages listed in the sources, read on 2026-10-05. Sume facts come from the Image API docs and the catalog code on main on the same date. Catalogs and limits change, so read the descriptors from GET /v1/images/models before you pin a number in production code.