Image reference limits: 10 on ElevenLabs, 16 on Sume, 5 Ideogram 4.5

ElevenLabs lists up to 10 reference images for gpt-image-2.5-flare and gpt-image-2.5-sunburst. Sume's Image API docs list up to 16 for both models on POST /v1/images, and 5 in total for Ideogram 4.5. The same model name carries different limits on different platforms, so a reference count is a property of the endpoint, not the model.

Check the limit where you call, and guard the count in code.

The limits side by side

As read on 2026-10-03.

Guard the count from the catalog

Sume publishes input_references as a range descriptor on each model, so the limit can be read at runtime instead of hard-coded. This function fetches it and trims or rejects before you spend a request.

import os, requests

H = {"Authorization": "Bearer " + os.environ["SUME_API_KEY"]}

def max_refs(model_id):
    r = requests.get("https://api.sume.com/v1/images/models", headers=H, timeout=60)
    r.raise_for_status()
    for m in r.json()["data"]:
        if m["id"] == model_id:
            d = m["supported_parameters"].get("input_references")
            return d["max"] if d else 0
    raise KeyError(model_id)

def guard(model_id, urls):
    limit = max_refs(model_id)
    if len(urls) > limit:
        raise ValueError(f"{model_id} accepts {limit} references, got {len(urls)}")
    return urls

# guard("openai/gpt-image-2.5", my_urls)

Rules for choosing references

More references is not better references.

• Order matters for Ideogram 4.5 on Sume: the first image is the one being edited and the rest are references.
• URLs must be public HTTPS. Localhost, private network and non-HTTPS addresses are rejected before submission.
• On edit calls, use aspect_ratio: "auto" to match the reference. Omitting the field is not the same, per the Sume docs.
• Use the fewest references that pin down the subject, style and product. Each one adds input cost on token-priced models.

Portability

If you maintain one prompt library across platforms, store a per-platform limit next to each model id and test your largest reference set against the smallest limit. Anything that works with 5 references on one endpoint will work on all three here; anything that needs 12 is only portable to the 16-limit endpoint.