Why input_references fails on Imagen, Recraft, Soul and Qwen Max

A reference-image request fails on imagen-4-fast, imagen-4-ultra, recraft-v4, higgsfield-soul and qwen-image-max because the catalog lists their input_references descriptor as {min: 0, max: 0}: they are text-to-image only. The Image API says such a model rejects references, and any parameter a model does not list returns 400 unsupported_parameter; Sume does not silently drop it.

Check before you send

Read the descriptor once and cache it. This prints each model with its reference ceiling.

import os, requests

r = requests.get(
    "https://api.sume.com/v1/images/models",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    timeout=30,
)
for m in r.json()["data"]:
    refs = m["supported_parameters"].get("input_references", {})
    print(m["id"], refs.get("max", 0))

Reference ceilings and prices

What to do instead

For image-to-image, pick a row with a non-zero ceiling. If you want the look of a text-only model applied to your photo, generate with that model and then use a reference-capable row to merge the two.

A defensive wrapper

Fail before you send, not after.

def can_reference(models: list[dict], model_id: str) -> bool:
    for m in models:
        if m["id"] == model_id:
            refs = m["supported_parameters"].get("input_references", {})
            return refs.get("max", 0) > 0
    return False

print(can_reference([{"id": "a", "supported_parameters": {"input_references": {"max": 0}}}], "a"))

The other limits are separate

The reference ceiling is its own descriptor, not the output count n. n is capped at 4 on most rows (1 on grok-image), so a gpt-image-2.5 edit can carry up to 16 references and still return up to 4 images. Rows such as ideogram-v4.5 stop at 5 references, and most other edit-capable rows (nano-banana-2.1, seedream-5-lite, flux-2-pro, qwen-image) allow 10. Read the live descriptor rather than hard-coding any of these.