Text-to-image-only models on Sume reject input_references

Not every image model on a router does image-to-image. In the Sume catalog, a model that cannot edit reports input_references as a range from 0 to 0. If you send a reference anyway, the API answers 400 unsupported_parameter with the message that the model is text-to-image only, rather than ignoring the image and spending money on a text-only render.

From the catalog data in the repo, five ids are text-only: higgsfield/soul, google/imagen-4-fast, google/imagen-4-ultra, recraft/recraft-v4 and qwen/qwen-image-max.

Text-only versus edit-capable

A guard

Check the descriptor, then fall back to an edit-capable id if you have references. This keeps a pipeline from failing on a model swap:

import os, requests
headers = {"Authorization": "Bearer " + os.environ["SUME_API_KEY"]}

def accepts_refs(model_id: str) -> bool:
    url = f"https://api.sume.com/v1/images/models/{model_id}/endpoints"
    r = requests.get(url, headers=headers, timeout=30)
    r.raise_for_status()
    p = r.json()["endpoints"][0]["supported_parameters"]
    return p.get("input_references", {}).get("max", 0) > 0

print(accepts_refs("google/imagen-4-ultra"))
print(accepts_refs("openai/gpt-image-2.5"))

Why this matters for trend models

New image releases often split into a text model and an edit model. The same split can appear in a catalog, so treat edit support as a per-id fact and read it from the descriptor every time rather than from a model's name.

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.