Video editing API or AI prompt edit: which Sume endpoint to call

If the change is mechanical, such as cutting a range, burning captions, dimming or cropping, call a deterministic Sume video endpoint. If the change is what is in the picture, such as swapping a product or changing the weather, call the Omni edit mode with a prompt. Mechanical jobs run ffmpeg and use no model inference, so they are the better first stop.

Which endpoint does what

The table lists what each endpoint does, where the source clip has to live, and whether a generative model is involved. The trim docs say there is no provider inference, only worker ffmpeg, and that each trim uses a flat number of credits; the video-frames docs say it charges by compute time.

Where the prompt edit stops

Edit by prompt has two limits to design around. The output follows the source clip's framing and length, so you cannot change the aspect ratio with it, and the audio is always native and synced. To change the framing, crop or generate fresh. To shorten, trim first and edit the trimmed clip, which also lowers the edit bill since Omni bills per second.

A sensible order

A common order is trim, then edit, then caption. Trim the shot you want, run one prompt edit on it, check the result by pulling frames at the start and end, and burn captions last so the text is not regenerated by the edit. Each step is a separate job with its own Idempotency-Key.

• Trim first, so you pay to edit only the seconds you keep.
• Edit once, at 360p if you are testing, then at the final resolution.
• Burn captions last.
• Use video-frames to compare the first and last frame before and after.