
The sample episode is worth opening alongside this page: the shot text, the take history and the regenerate controls are all visible on it.
Regenerating an episode's storyboard rebuilds that episode from scratch, and every video you have already paid to generate in it is discarded with the old segments. The credits are not returned — they were spent on clips that were delivered. The app makes you confirm, and refuses outright if a video in that episode is still generating.
Regenerating one segment's video is a different thing and is safe: the old take stays in that segment's history and switching back is free. The mechanics of both are covered in fixing one shot without redoing everything. This page is about what to change when a take comes out wrong, and what each change costs.
Free: editing any of the storyboard text, changing which setup image a segment uses, switching between takes you have already generated, previewing, and re-reading anything you already have.
Charged: generating video. Each attempt is a separate model call and is billed as one. There is no “try again free” — that is the honest shape of a service that pays per call upstream.
Which is why the order matters. Fix the text first, look at it, and only then spend. A cheap tier is a reasonable place to find out whether a shot reads before you spend a good one on it.
Most bad takes trace back to something in the shot text rather than to the model having a bad day. The table below is the short version; each fix is an edit to the segment, which costs nothing until you press generate again.

The video draws what a shot shows, and small things slip. Make the action something the camera can see: give it its own shot, frame it close — a close-up on her hand turning the key — and give it a second or two. Name the prop with @ in that shot and keep it on the Props line. Something worn on the body, such as a ring, a badge or an earpiece, reads far better in a close shot than in a wide one. A mechanism, such as a lock opening or a case being loaded, reads best as one clear motion rather than a list of steps. Then regenerate that segment. Props and locations in a shot.
When the clip is right apart from one thing in the picture — a colour, the weather, an expression — Edit this clip under the player may fix it without regenerating; prices are on the pricing page. Edit this clip.
Stop is not cancel-the-order. If the job has already been picked up and sent upstream, stopping it releases the interface but the credits are settled as spent — the call was made and we were billed for it.
If the job has not been picked up yet — still queued, or parked because your slots were full — stopping it does refund, in full, because nothing was ever sent.
A segment starts saying “generating” the moment you submit it, which is before anything has been sent — so a segment showing that is often still fully refundable, and there is no way to tell from the card. If you have changed your mind, press stop and find out. Once it really has gone upstream, stopping costs the same as letting it finish, and letting it finish at least gives you a take.
Automatic: the model never received a usable request; the model answered that it was busy or rate-limited; a clip was refused by content moderation.
Moderation refunds have brakes on them, because an unlimited auto-refund on refused content is an invitation to generate refused content all day. There is a per-account limit within a 24-hour window, a site-wide daily ceiling, and a master switch. The part worth knowing: crossing the per-account limit does not just stop the refund — it suspends your video generation for an hour, and the counter does not reset when the suspension lifts, so each further rejection suspends you again.
Not automatic: a timeout or a dropped connection mid-call. From our side those look identical whether the model ran or not, so the credits stay frozen — neither charged nor returned — and a person settles it. The same applies if the model produced a clip but we failed to deliver it to you: the amount stays frozen rather than being charged, because you did not receive anything.
In all of those cases the credits panel in the top bar shows the ledger, and the line will say which of the three happened.
Generation is probabilistic. The same text can give a good take and a bad one, and after two or three attempts with no change in the text you are sampling, not fixing.
The signal that editing will help is that you can point at a sentence and say what it is telling the model to do wrong. If you cannot, the text is probably fine and the shot is a dice roll — in which case the cheapest move is to generate again, keep the better take, and move on. Both takes stay in the history either way.
| Symptom | Usually caused by | Fix | Cost |
|---|---|---|---|
| A person appears who should not be in the shot | The shot's in-frame list includes them — that list, not the prose, decides who gets a face | Edit the shot so they are described as off-screen, or remove the mention | Free until you regenerate |
| The wrong character's face | Two characters described too similarly, or a missing setup image | Give each a distinct look in the asset library, generate the setup image | Setup image is charged |
| Wrong outfit or wrong state (before/after an event) | The segment is using the base image rather than a state image | Generate the state image and pick it for that segment | State image is charged |
| Framing repeats the previous segment | The previous segment's last frame is attached as a reference for continuity | Edit the previous segment's text — that unhooks its frame. Regenerating alone re-uses it | Editing is free; the regeneration is charged |
| Background crowd appears and disappears between segments | Only some segments declare background people | Add or remove the background-people line consistently across those segments | Free until you regenerate |
| Segment is longer or shorter than expected | The shot durations in the text sum to the segment length | Edit the per-shot seconds | Free until you regenerate |
| Lines are spoken by the wrong voice | The voice tag on the line | Edit the tag, or reassign the voice | Free |
No. Each take is kept in that segment's history and you can switch back to any of them for free. The destructive action is regenerating the whole episode's storyboard, which discards every segment in it.
Yes. Text edits, image choices and version switches cost nothing. Only generating video is charged.
Only if the job had not started yet. Once it has been sent upstream, stopping releases the interface but the credits are settled as spent, because the call was made.
Three possibilities: refunded automatically (the model never ran, was busy, or refused on moderation), charged (it ran and delivered), or still frozen pending a manual check (a timeout, or a delivery failure on our side). The ledger line says which.
Because it would discard the segment you are currently paying to generate. That one has no override.
If you cannot point at the sentence causing the problem, two or three. Past that you are sampling randomness rather than fixing anything, and a cheaper tier or a rewritten shot is a better use of the credits.
Editing costs nothing and it is where most bad takes actually come from. Generate only once you have read the shot back and it says what you meant.
Updated October 2, 2026
Pricing·Novel to Video·Guides·Samples·Support·Privacy·Terms·Legal·Contact·© 2026 SceneMixer