Nano Banana 2 image search grounding: what Sume lets you send instead

Google's image generation docs say Nano Banana 2 (gemini-3.1-flash-image) integrates both Web Search and Image Search, and that Image Search can be used alone or together with Web Search (Google AI for Developers, read 2026-10-04). The same page lists up to 14 reference images in total. If you call Nano Banana 2 through Sume, you will look for the grounding switch and not find one.

What Sume exposes

Sume's image catalog publishes typed descriptors for each model: prompt, aspect ratio, resolution, quality, n, output format and input_references. A model accepts only the parameters it lists, and anything else returns 400 unsupported_parameter (Sume Image API). None of those descriptors is a search or grounding field, so there is nothing to switch on.

The workaround: be the search step

Grounding exists so the model can see what a real thing looks like. You can do that step yourself. Find the reference photos you are allowed to use, host them at public HTTPS URLs, and pass them in input_references. Say in the prompt what each image is. You decide which pictures count, and the job record shows exactly what was sent.

• Use only images you have the right to use as references.
• Keep to 10 references on Nano Banana 2 through Sume, per the docs' general rule of 10 (16 on ChatGPT Image 2.5).
• Log the reference URLs with the job, so a result can be traced to its inputs.

A request with fetched references

Post this body to POST /v1/images with your bearer key.

{
  "model": "google/nano-banana-2",
  "prompt": "A product shot of the object in Image 1, set in the kind of kitchen shown in Image 2. Keep the object's shape and label exactly.",
  "input_references": [
    {"type": "image_url", "image_url": {"url": "https://example.com/object.png"}},
    {"type": "image_url", "image_url": {"url": "https://example.com/kitchen.png"}}
  ],
  "aspect_ratio": "auto"
}