Nova Canvas returned fewer images than asked: Sume n and data length

Ask Nova Canvas for four images and you may get three. Amazon's docs (read 2026-10-06) say numberOfImages can return fewer images when responsible-AI moderation blocks some of them. Code that assumes four will index past the end of the list.

The Sume equivalent

n on Sume works the same way from the caller's side: you ask for a count, the response has a data array, and the per-model maximum comes from the catalog. The safe pattern is to never assume len(data) == n. Count the array and divide the billed amount by the count when you track cost per image.

usage.cost is the billed USD amount. A job that fails terminally returns a 502 error envelope with category, retryable and next_action, and failed jobs are not billed. For a slow request the response is a 202 job envelope, and the images come from the result URL, not from the first response.

Count first, then price

The function below handles three outcomes: a full result, a short result and a job that is still running.

import os, requests

def edit(n: int):
    r = requests.post(
        "https://api.sume.com/v1/images",
        headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
        json={"model": "google/nano-banana-2", "n": n,
              "prompt": "Same mug, new table settings.",
              "input_references": [{"type": "image_url", "image_url": {"url": "https://example.com/mug.jpg"}}]},
        timeout=90)
    if r.status_code == 202:
        return None, r.json()
    r.raise_for_status()
    body = r.json()
    got = len(body["data"])
    per_image = body["usage"]["cost"] / got if got else 0
    return got, per_image

print(edit(4))

What to log

Log requested n, returned count and usage.cost on every call. A gap between the first two on a prompt that keeps coming back short usually points at the prompt, and it is cheaper to rewrite the prompt than to retry.