Guides

    Gemini Omni Flash Edit: retouch the take, do not re-roll the idea (13cr)

    Gemini Omni Flash Edit edits footage you already have. A new generate is a different identity.

    Versely Team4 min read

    Gemini Omni Flash Edit edits footage you already have. A new generate is a different identity.

    That is the whole product. Google files this row as conversational video editing — video-to-video — in the edit-video family. It does not invent a shot. It revises a file. The catalog lists 13 credits, no native audio, no duration enum, no output resolution. The input is a video, not a still: requires_image is false because the source is video_url.

    If the plate is already the idea, stay on Gemini Omni Flash Edit. If the idea is still a sentence, you are on a generate row, and you will spend 13 credits a second to lottery a cousin of the take you meant to keep.

    The take is the identity

    A take is the blocking, the hand, the label on the pack, the light the client already said yes to. Edit is the job that treats those pixels as ground truth. You write the change: warmer grade, red jacket, tighter window, mug swapped for the branded can. Everything you did not name is supposed to survive.

    Re-roll is a new sample. Same prompt, new person, new table, new logo that is almost the logo. Teams do this because the bill looks the same — 13 credits a second on the Omni Flash family — and because "one more generate" feels faster than writing a precise note. It is not faster. You now have two almosts, and the approved identity is gone.

    The agent path that means this is edit or redo a previous generation. Redo is allowed when the idea was wrong. Edit is required when the idea was right and the pixels need a note.

    What this endpoint will not invent

    Audio. Catalog audio is false. Do not ask the mouth to say a new line and call the result a dialogue fix. Lay a voice pass after the picture is right, or generate a talking shot that was always meant to speak.

    A second camera. One continuous plate in, one revised plate out. Three setups is three generates, then a cut in video editing. Chaining edit turns will not grow extra lenses.

    A blank start. There is no clip yet? That is text-to-video or image-to-video. Edit requires the file. Google's Omni Flash generate rows are a different identity problem; this page is only the retouch.

    The longer loop — turn by turn, what to type — is conversational video editing with Gemini Omni Flash. The routing rule on this page is narrower: if you would be sad to lose this exact take, you are not allowed to generate your way out of a retouch.

    Thirteen credits, scoped to the source

    Billing is per second on the source duration. The catalog floor sits at 63 credits and the ceiling at 1,500, because long plates are long bills. That is an argument for a short, locked clip and a small instruction, not for a fresh generate "to save a pass." A generate re-decides the identity for the same rate. An edit spends the rate on the change.

    Write the note so it could only apply to this file. "Change the jacket to navy, keep the walk and the kitchen." If the sentence could apply to any kitchen, you are describing a new shot. Send that to generate. Keep this row for the file you already have.

    Google's roster puts Omni Flash next to Veo. Veo is how you start a cinematic clip. Omni Flash Edit is how you keep one.

    FAQ

    Can I use Omni Flash Edit when I only have a still?

    No. This row wants a video in. A still is an image-to-video job, then an edit if the motion is already a keeper. Do not paste a jpg into an edit endpoint and hope it becomes a take.

    If both edit and generate cost 13 credits a second, why not always generate?

    Because generate samples a new identity. You pay the same rate for a different person. Edit pays that rate to change the plate you are keeping. The expensive version of a retouch is the afternoon of "almost the same shot" retries, not the sticker on the first call.

    Will the soundtrack come back rewritten?

    Treat the endpoint as a visual pass. Catalog audio is off. If the VO has to change, do that as a separate speech or dub step after the picture lock.

    When is a redo actually the right call?

    When the blocking is wrong, the room is wrong, or you would not miss this take. Then generate. The agent can redo the previous generation without pretending the new file is the old one.