Omni Flash conversational edit is not a generate
Talking the timeline into a change is an edit job. A new shot is a generate job. Do not pay generate prices to fix a crop.
A crop is not a new shot. A jacket colour is not a new shot. A warmer grade on a plate the client already approved is not a new shot. Talking those changes into the clip is an edit. Asking the model for another eight seconds from a blank prompt is a generate. They are different endpoints, different failure modes, and a different way to waste credits.
Gemini Omni Flash Edit is the edit job. It takes an existing video plus an instruction and returns a revised clip. The generate family — text-to-video, image-to-video, a new sample from a prompt — is how you start a shot. Do not start a shot to fix a crop.
The loop itself, turn by turn, is conversational video editing with Gemini Omni Flash. This page is only the routing rule.
Two prices that look alike and are not
On Versely both Omni Flash text-to-video and Omni Flash Edit bill at 13 credits a second. That is how people talk themselves into generating again. The rate is the same, so "just reroll it" feels free relative to editing.
It is not. A generate is a new sample. Identity, blocking, the hand position, the label, the light the client signed — all of that goes back into the lottery. The second generate that "almost" matches the first is the one you then have to edit anyway, except now you are editing a cousin of the approved plate. An edit call is billed on the source duration and is scoped to the change. The camera move you already liked is not re-decided.
Fixing a crop with a generate is paying for a new film to move a window. Fixing a crop with an edit is moving the window.
Use the video editor when the change is a hard cut, a title, a resize that does not need new pixels, or a sequence of shots that were never one clip. Use Omni Flash Edit when the change is inside one continuous plate and the plate is worth keeping.
What belongs on Edit
One object, one attribute. Replace the mug, change the jacket, swap the box for the branded pack. The instruction names the thing and says everything else unchanged.
A restyle that should keep the blocking. "Golden hour, same framing." "Shot on 16mm, same move." The motion is the reason the restyle looks like a grade instead of a different video.
A crop, a reframing, a cleanup. This is the one people generate for, and it is the one they should not. The source clip already has the performance. You are asking for a different window onto it, or for a defect to leave, not for a new performance.
Chain those. Edit two sees the output of edit one. "Make the jacket red" then "now make it evening" is one plate with two notes. Two generates with those prompts is two coin flips.
What belongs on generate
There is no plate yet. You do not have a clip. You have a sentence. That is text-to-video, or a still then image-to-video, not Edit. Edit requires a video in.
The blocking is wrong. A new angle, a new room, a different performance. Edit cannot recompose a shot from a camera that was never in the file. If the idea is wrong, regenerate. If the last frame is good and you want more of the same move, that is extension, which is also not Edit.
The line has to change. Versely's Edit endpoint is a visual edit. It does not resynthesize a spoken line inside the clip. A new line is a voice pass, a dub, or a generate that was always a talking shot. Do not spend an edit turn asking the mouth to say something else and calling the result a dialogue fix.
Three different setups. Omni Flash Edit operates inside one continuous clip. A sequence that needs three camera positions needs three generates (or three locked stills animated), then a cut in the editor. Three edit turns will not invent the other two cameras.
A cheap test before you spend
Write the note as a sentence. If the sentence starts with "change," "replace," "crop," "warm," "remove," or "keep this and…," you are on Edit. If it starts with "a shot of," "new," "another take," or "from a different angle," you are on generate.
Then look at the plate. If you would be sad to lose this exact hand, this exact label, this exact move, you are not allowed to generate to save a few seconds of thinking. Edit the plate. If you would not miss this take, generate. That is the whole rule.
Credits follow the rule rather than the other way around. The expensive version of a crop is not the 13-credit second. It is the afternoon of identity retries you bought because you treated an edit as a generate.
FAQ
Can I use Omni Flash Edit to recrop a shot instead of generating a tighter take?
Yes. That is one of the jobs the endpoint exists for. The source clip keeps the performance; the instruction asks for a different window. Generating a tighter take is how you lose the performance you already had.
Why not always generate, if both cost 13 credits a second?
Because a generate re-samples the scene. You pay the same rate for a new identity. An edit pays that rate to change a plate you are keeping. The retry tax on "almost the same shot" is what makes generate-to-fix expensive, not the sticker on the first call.
Will an edit change the audio?
Treat Versely's Edit endpoint as a visual pass. Do not use it to rewrite a spoken line. If the VO has to change, do that as a separate voice or dub step after the picture is right, or generate a talking shot that was always meant to say the new line.
When should I stop editing and cut in a real editor?
When you need a hard cut, titles, a second clip, or precise audio sync across a sequence. Conversational edit is for narrowing one plate. Video editing is for assembling plates. Past a handful of turns on a long clip, the conventional timeline is cheaper than another edit call on the whole duration.