Runway Ad Localization recipe vs a Sume image edit prompt

Runway's Ad Localization Recipe takes a referenceImage and a targetLanguage code such as ja or es and returns the same ad with its on-screen text adapted. Sume has no language-code field. Send the ad as a reference image to POST /v1/images and write the language, and ideally the translated lines, into the prompt, one language per call.

Recipe facts are from Runway's page; Sume facts from the Image API docs, both read 2026-10-01.

What does the Runway recipe ask for?

Three inputs: a pinned version, the referenceImage as { uri }, and targetLanguage. The page says codes are ISO-style, so use ja, not "Japanese". Its tips are to start from a high-resolution ad with readable text, and note that clear typography localizes more reliably than heavily stylized layouts.

How do I do it on Sume?

Put the ad in input_references as a public HTTPS image URL; private-network and non-HTTPS URLs are rejected before submission. Set aspect_ratio: "auto" so the output matches the reference, since the docs say omitting the field is not the same as auto. Then write a prompt such as: keep the layout, photo and logo unchanged, and replace the headline and button text with this Spanish copy, followed by the exact lines.

How do I run many languages safely?

Submit one job per language and give each its own Idempotency-Key, for example ad-17-es. A replay returns the original job, so a retry after a timeout does not pay twice. Do not reuse a key for another language: that is an idempotency_conflict. Check the queue limits in Generation admission before sending a big batch.

Should I check the output text?

Yes. A prompt is not a translation table, and Sume's docs do not promise exact text rendering, so read every localized image before it ships. The post on exact copy mapping covers keeping the lines fixed.