Fix the last second of an Omni clip: edit it or generate a new ending

If only the last second of a 10-second Omni clip is wrong, the cheapest fix is to trim the good part, take its last frame, and generate a new 3-second ending from that frame: about $0.38 at 720p plus a $0.02 trim. A video edit of the whole clip costs $1.25 at 720p, and re-renders everything.

Both routes run on gemini-omni-flash-1.1 through the Sume Video Router, whose edit mode takes video_url and a prompt, and whose image-to-video mode takes image_url (Video Router page).

Route A: edit

Send the clip as video_url with a prompt that names the fault: "At the end, the cup is held in the right hand, not the left. Keep everything else the same." The source sets the output length, and no aspect_ratio or duration is sent. Cost is the output seconds at your chosen resolution, which defaults to 720p. Because the whole clip is regenerated, the first nine seconds can change too, which is the main risk.

Route B: trim, last frame, new ending

Cut the keeper with the Video Trim API ($0.02 per job). Extract its last frame with the Video frames API. Send that still as image_url and generate a 3 to 4 second ending; then join the two on a Timeline. The keeper stays untouched, so you only pay for the new seconds.

Request for route B

The new ending starts from a still, so describe the motion you want.

{
  "model": "gemini-omni-flash-1.1",
  "prompt": "The hand lifts the cup toward the camera and smiles, same light, steady shot",
  "image_url": "https://media.sume.com/artifacts/artf_demo/last-frame.jpg",
  "resolution": "720p",
  "duration": 3,
  "aspect_ratio": "9:16"
}

Pick by the risk

• The first nine seconds are perfect and you must not touch them: route B.
• The fault is small and spread through the clip (a color, an object): route A.
• The ending needs the exact motion of the original: route A, since a still has no motion.
• Sound matters: every Omni clip has native audio, so route B has a join in the sound. Lay one audio spine under the join, or check it by ear.

Before you pay

Run route B at 360p first: a 3-second draft ending is $0.12. It tells you whether the ending prompt works, and the keeper is not touched. If the draft is right, repeat at your delivery resolution. Route A can also be tried cheaply, but an edit at lower resolution gives you a lower-resolution clip, so use it to test the prompt rather than the picture.

A worked example

Suppose a 10-second 9:16 clip of a barista is perfect until second 9, when the cup morphs into a different shape. Route A, the edit, sends the clip with "keep everything the same, the cup stays a white ceramic cup in the right hand" at 720p: $1.25, and the first nine seconds may change slightly. Route B trims seconds 0 to 9 ($0.02), takes the frame at 8.9 s, and generates a 3-second ending at 720p ($0.38). The total is $0.40 plus joining.

Run route B first at 360p ($0.12) to check the ending; that makes the whole experiment $0.14. If you need to try three endings, it is $0.42 at 360p against $3.75 for three whole-clip edits at 720p.

What to watch at the seam

Look at the join frame by frame: position of the subject, direction of motion and light. A still carries no motion, so the new ending can start slower or faster than the keeper. Say in the prompt that the motion continues in the same direction. If the sound jumps at the join, lay a single audio track over the joined clip.

Join costs

Route B adds assembly work. The Timeline API takes ordered video slots over one audio spine and returns one MP4, with inputs that are already media in your workspace. Compare the extra step with the $0.87 you save at 720p ($1.25 less $0.38) before you decide that route B is worth it.