Multi-reference image prompts when the API has no role field

In Ideogram's app, you can mention attached images inline with @ syntax, according to its reference image docs. Google's Gemini API image guide goes further on Nano Banana 2 and splits its 14 references into object images and character images.

Sume's request shape has neither. An input_references entry is {type: "image_url", image_url: {url}}, and nothing else. There is no role field, no weight and no tag. The one ordering rule the docs state is for Ideogram 4.5, where the first image is the one edited.

Say it in the prompt

Since the API has no field for roles, put the role in the sentence. Refer to images by order and by what they show, and keep each instruction about one image.

A request

import os, requests
headers = {"Authorization": "Bearer " + os.environ["SUME_API_KEY"]}
urls = ["https://example.com/product.jpg", "https://example.com/style.jpg"]
body = {
    "model": "openai/gpt-image-2.5",
    "prompt": "Place the product from the first image in the setting and lighting of the second image. Keep the product unchanged.",
    "input_references": [{"type": "image_url", "image_url": {"url": u}} for u in urls],
    "quality": "medium",
}
r = requests.post("https://api.sume.com/v1/images", headers=headers, json=body, timeout=60)
print(r.status_code)

Test the order

How a model weighs image order is model behavior that the Sume docs do not specify, so run the same prompt with the references swapped once and keep the order that works. Keep a short list of working prompt patterns per model; they transfer less than you would expect.

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.