Remove a person or object from a finished clip with Gemini Omni edit

To remove a person or object from a finished video with Gemini Omni Flash 1.1, send the clip as video_url with a prompt in the form "Remove X. Keep everything else the same." There is no mask or region field on Sume's edit request, so the prompt is the only way to say what goes. The output length follows the source clip, and a 720p result costs $0.125 per second.

The request

Sume's edit mode accepts video_url plus a prompt, and optionally resolution (default 720p). It does not accept aspect_ratio or duration as creative controls, and video_url cannot be combined with image or reference fields. The docs example prompt, "Replace the bottle with an apple. Keep everything else the same.", is the pattern to follow with the verb swapped to remove.

import asyncio, json, os, urllib.request

def post(path, body, key):
    req = urllib.request.Request(
        "https://api.sume.com" + path,
        data=json.dumps(body).encode(),
        headers={
            "Authorization": "Bearer " + os.environ["SUME_API_KEY"],
            "Content-Type": "application/json",
            "Idempotency-Key": key,
        },
    )
    with urllib.request.urlopen(req) as resp:
        return json.load(resp)

async def main():
    out = await asyncio.to_thread(post, "/v1/video-router/generate", {
        "model": "gemini-omni-flash-1.1",
        "video_url": "https://media.sume.com/example/street.mp4",
        "prompt": "Remove the person in the red coat. Keep everything else the same.",
        "resolution": "720p",
        "mode": "async",
    }, "remove-person-001")
    print(json.dumps(out, indent=2))

asyncio.run(main())

Write the prompt like a work order

Name the thing by what it looks like and where it is, not by what it is. "The person in the red coat on the left" is easier to find in the frames than "the extra".

• One removal per request; chain a second pass for a second item.
• Say what should fill the space only if it is not obvious, for example "the brick wall continues behind".
• End with the keep clause, so the model treats everything else as fixed.
• If the object is only in part of the clip, say when: "until the 4-second mark".

What the edit costs

Billing for an edit is reserved on a duration hint (default 8 seconds, ceiling 30) and the output follows the source length, so check the final charge against the clip you actually sent. For a 10-second clip at the default resolution the rate table gives $1.25.

When to cut instead of edit

If the unwanted item appears only in the first or last seconds, trimming is cheaper and safer than regenerating anything. Video trim costs $0.02 per job and takes start plus end or duration. Use the edit only when the object stays in frame for most of the clip, or when a cut would break the shot.

Edit output follows the source length, and the reserve is a hint that tops out at 30 seconds, so trim long sources to the part that needs the change before you send them.

Check what was removed, and what else moved

Removal is the edit most likely to leave a trace: a shadow, a reflection in glass, or a hand still holding a missing cup. Pull stills from the result at the moments the object was visible and look at the places around it. A frame-pair check with video-frames is cheaper than watching the clip twice.

A second concern is the rest of the scene. The prompt says to keep it, but the edit regenerates the clip, so faces, text on signs and fine patterns can change. If those matter, compare them against the source before you publish. See also swapping the product in a finished video for the replace variant, and the reserve and 30-second cap for the billing detail.