Nano Banana 2 Lite takes 14 references; Sume's cap is 10

Google's image generation docs (read 2026-10-05) allow up to 14 reference images in total on gemini-3.1-flash-lite-image (Nano Banana 2 Lite). Sume's Image API caps input_references at 10 on catalog models and 16 on ChatGPT Image 2.5. Lite is not in the Sume image catalog today, but google/nano-banana-2 is, so a 14-image set means trimming to 10 on Sume. Check GET /v1/images/models for the live number.

Reference limits by model

The Google figures are from its docs, read 2026-10-05. The Sume figure is from the Image API docs.

Two differences to plan for

First, Google counts Lite's 14 as one pool, while NB2 and Pro split their slots by role (object, character, style). Sume's input_references is one flat list, and we make no claim that Sume maps Google's slots. Second, the count is higher on Lite than on Sume's Nano Banana 2 row, so a set built for Lite will not fit.

• Drop duplicates first: two near-identical product shots use two slots.
• Merge small items into a contact sheet: a 2 x 2 image of four swatches uses one slot.
• Keep the identity references. Remove ambient mood images before anything that defines the subject.
• If you must send more than 10, ChatGPT Image 2.5 takes 16 on Sume.

Trim to ten

This sends the first 10 of your references to Sume's Nano Banana 2 row. Put your most important references first.

import asyncio, os, httpx

REFS = [f"https://example.com/ref-{i}.jpg" for i in range(14)]

async def main():
    body = {
        "model": "google/nano-banana-2",
        "prompt": "the same character in a rainy street, cinematic",
        "input_references": [{"type": "image_url", "image_url": {"url": u}} for u in REFS[:10]],
    }
    headers = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
    async with httpx.AsyncClient(timeout=60) as c:
        r = await c.post("https://api.sume.com/v1/images", headers=headers, json=body)
    print(r.status_code, r.text[:200])

asyncio.run(main())

Related reading

For the Nano Banana family more broadly, see the earlier 14-reference comparison. Sume's catalog price for google/nano-banana-2 is on the endpoints call, and Google's own per-image pricing is on its pricing page.

A fair way to test

Run the same prompt with 14, 10 and 6 references, if your tool allows it, and see whether the extra four change anything you care about. If they do not, the Sume cap costs you nothing. If they do, note which four mattered; they are the ones to keep when you trim.