Runway Ad Localization in 20 languages vs a Sume bulk run

Runway's Ad Localization app takes one ad image and one target language and returns a localized image that keeps the layout, at up to 21 credits per generation, with 20 supported languages listed. At the 21-credit ceiling, 19 languages from an English source is 19 x 21 = 399 credits, and all 20 would be 420. On Sume you build the same set as a bulk queue: one run per language, each with the ad image attached and the language named in the instruction.

What the Runway app does

The Sume version

A bulk request is a server-side queue of ordinary Format runs. The envelope has two keys: concurrency, an integer from 1 to 16, and items, 1 to 100 entries, each the same body as a single run. Each item can carry its own attachments, instruction and webhook. Mint a fresh Idempotency-Key per batch, because replaying a spent key returns the old queue.

The script below builds 19 items for a Format you own, here called acme/ad-localize, and prints the body size. It sends nothing.

import json

LANGS = ["Arabic", "Chinese (Simplified)", "Chinese (Traditional)", "Dutch", "French",
         "German", "Hindi", "Indonesian", "Italian", "Japanese", "Korean", "Portuguese",
         "Russian", "Spanish", "Swedish", "Thai", "Turkish", "Ukrainian", "Vietnamese"]
AD = "https://cdn.example.com/ads/spring.png"

items = [
    {
        "instruction": f"Translate all visible text to {lang}. Keep layout and style.",
        "attachments": [{"type": "input_image", "image_url": AD}],
        "generation_spend_cap_usd": 3,
    }
    for lang in LANGS
]
body = {"concurrency": 4, "items": items}
assert 1 <= len(items) <= 100
print(len(items), "items;", len(json.dumps(body)), "bytes")
# POST to /v1/formats/acme/ad-localize/bulk-runs with an Idempotency-Key header.

Read the queue honestly

The queue status completed means every item is terminal, not that all succeeded: branch on counts.failed. The queue has no webhook, so poll GET /v1/format-run-queues/{id} or register a webhook per item.

Before you scale a batch

Run one item first and read the receipt, including usage and the spend cap, then queue the rest. A bulk queue is complete when every item is terminal, so check counts.failed rather than assuming success.