Meta Feed allows 1-second video ads: shortest Sume clip is 2 s

Meta's Facebook Feed video spec, read on 2026-10-07, allows a video length from 1 second up to 241 minutes. No Sume video model generates a 1 second clip: the shortest duration in the catalog code is 2 seconds on Wan 3.0, then 3 seconds on Gemini Omni Flash 1.1, then 4 seconds on the Seedance, Kling and Grok models and 5 seconds on MiniMax H3. If you want a true one-second cut, generate the shortest clip that fits and trim it down with the video-trim route.

What each model's floor is

The table is from the capability records behind GET /v1/videos/models. Durations are integers, and the catalog lists every whole second between the minimum and the maximum.

Why you might want one second

A one second ad is a test of the very first frame. In a feed, the scroll decision is made fast, and a short cut of only the opening tells you whether the first image earns attention at all. It is not a finished ad. Treat it as a way to compare openings cheaply, then build the full cut only around the opening that wins.

This is also where a hook test gets confusing. If two arms differ in more than the first second, you cannot tell what mattered. Cutting every arm to the same short length removes the length and everything after the opening from the comparison.

How to cut it

Generate the opening at the shortest duration your chosen model allows, then call POST /v1/video-trim with the artifact URL, a start of 0 and a duration. The trim takes video_url, start and exactly one of end or duration, requires an Idempotency-Key, and costs $0.02 per job, so a hundred one-second openings cost $2.00 in trim jobs. The result is a new artifact and the source is never overwritten.

The docs do not state a minimum output length for trim, so check the returned duration_seconds and any warnings on the first job before you batch.

curl -sS -X POST https://api.sume.com/v1/video-trim \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: hookA-open-1s-v1" \
  -d '{"video_url":"https://media.sume.com/artifacts/artf_demo/hookA.mp4","start":0,"duration":1}'

Cost of a trim-based opening test

Take ten openings. If each is generated at 4 seconds on a Seedance model, you pay for ten generations at that model's rate, which you read from usage.cost on the first completed job. The trims add 10 times $0.02, which is $0.20. The trim does not make generation cheaper. What it buys is a uniform, shorter file for everything downstream: captions, review and the ad platform upload.

• Generate at the shortest allowed duration.
• Trim every arm to the same length.
• Compare openings first; only then build full cuts.

Limits worth stating

Meta's page describes the allowed range, not the length that performs. We make no claim that one second ads are effective, and the spec may change. If your placement is not Feed, check its own spec, since the page we read did not give numbers for other placements.

A plan for a 10-opening test

Write ten one-sentence openings and give each a key such as open-01-v1. Submit ten bodies that differ only in the prompt, on one pinned model at its shortest duration. When the jobs complete, trim each to the same duration and keep the returned artifact URL next to the key. Then look at all ten as stills and as clips before you decide which three get a full cut.

A bulk Format queue is another way to run ten at once, but for a direct model call ten separate /v1/videos posts are simple, and each one carries its own idempotency key so a retry cannot double-bill. Keep the batch inside what your balance can reserve, since each submit reserves the list price times 1.25 for that model.