Amazon Sponsored Brands video: cut a 20-second window with Sume trim

If your product video runs 90 seconds and Amazon Sponsored Brands wants 6 to 45 seconds, cut it with POST /v1/video-trim and send start plus duration: 20. Amazon's spec page says the length is "between 6 and 45 seconds (20 secs or less highly recommended)", so a 20-second window is the safe target. Sume's video trim returns a new MP4 with only that range and leaves the source untouched. Public price is $0.02 per job.

The numbers on both sides

The trim limits come from the Sume docs, and the Amazon rows come from the vendor page.

Steps

Import the long clip with POST /v1/media-imports. The trim API does not fetch from the open internet and answers unsupported_media_source for off-host URLs.

• Pick the 20-second window you want to keep and note its start in seconds.
• Send start and only one of end or duration. Sending both returns video_trim_range_conflict.
• Leave precision on exact so the cut lands on the frame you chose. keyframe is a stream copy that can start up to a GOP early, so re-base your times on actual_start_seconds.
• Add Idempotency-Key, then poll GET /v1/jobs/:id/status and read the result for the new video_url.
• Use that URL as a slot in a Timeline render if you also need 16:9 output at 1920x1080.

The exact path also takes an output object with width, height and fps, so one trim can conform the size in the same job. Width and height run 256 to 2160, and fps is 24, 25, 30 or 60.

Choosing the window

A trim only helps if the 20 seconds are the right ones. For a holiday product video, the first three seconds should show the product, and the last two should carry the offer or the brand. A simple method is to pull eight stills with video_inspect, which is the default when you do not send frames, and pick the start where the product first appears clearly. Since inspect stills are mid-bin samples across the clip, note that a still at a given time only shows that instant.

If the clip has speech, a sentence-level transcript from inspect gives you natural cut points, so you do not end in the middle of a word. The transcript costs $0.01 per audio minute according to the docs, plus the inspect compute, and returns segments[] when you request segmentation.mode: "sentence".

Once you have a start time, the trim call is deterministic: the same start and duration give the same range, and the response reports actual_start_seconds so you can confirm where the cut landed.

What Sume does not do

Trim does not choose the best 20 seconds for you. It cuts the range you give it. Use video inspect stills if you want to look at candidate frames first. It also does not enforce Amazon's 500MB limit.