Why Does an AI Video Come Out Wrong?

Seven causes explain most failed shots. Text that has to be read, two actions in one sentence, a camera move that fights the scene, a face that drifts between shots, the wrong aspect ratio, the wrong tier for the shot, and a prompt never tested as a still. Every reshoot charges for the whole clip again — 490,000 tokens for a five-second draft at the base rate.

Why is a reshoot full price?

The token table charges per generation. There's no partial price for a missed shot and no refund for a deleted one, so a clip you throw away costs exactly what the one you keep does. 490,000 tokens on Draft, 665,000 on Standard, 700,000 for four seconds on Cinema, and 1,872,500 on Cinema Pro. Pro plans pay 20 percent less per transaction, the free plan 30 percent more.

That's why every fix below stops before the Generate button. A 29,750-token draft image answers most questions at a sixteenth of the cheapest clip.

The seven causes

1. Text to be read in frame. Text-to-video models break type and fine detail — the documented examples are warped hands and unreadable signs. Prices, labels, and logos are kept out of the generation and added at render on Cinema Studio's text-and-graphics lane, and the token table lists no charge for that.
2. Two actions in one sentence. A five-second clip carries one verb. A prompt with a character who enters and then sits produces one of them, and rarely the one you wanted. Split it into two scenes. Two five-second draft clips are 980,000 tokens, the same as one ten-second clip.
3. A camera move that fights the scene. One move from a fixed list of thirteen is applied per clip, and the move is a request the engine weighs against the scene description. A handheld tag on a static product shot, or an orbit around a subject moving away, ends in a compromise. Match each move to the shot it belongs to.
4. A face that drifts between shots. Every generation starts from scratch, so two prompts describing the same person in the same words produce two different people. Attach a character from the Characters tab, or open the second shot from the same still with Pick from library. A full character guide is 1,377,000 tokens; regenerating one wrong face once is 665,000 at the standard tier, and the character-consistency guide works out when the guide pays for itself.
5. The wrong aspect ratio. The ratio is baked into the shot at generation. Cropping a 16:9 clip to 9:16 keeps 31 percent of the frame width and shifts the subject off center. The target is set before the first clip, because the fix is regenerating every clip.
6. The wrong tier for the shot. Tiers run different engines, and the tier is chosen per clip. An atmosphere written for Cinema — asking for shallow focus and film light — comes out flat when generated on Draft. A two-second insert generated on Cinema Pro at 1,872,500 tokens looks fine and costs nearly four draft clips. Match the tier to the shot.
7. A prompt never tested. The sentence went straight to video. A 29,750-token draft image shows the subject, the light, and the framing in under a minute. Three stills are 89,250 — less than a fifth of a draft clip — and they catch causes one, four, five, and six before you pay the video price. What goes into that sentence is in the how to write an AI video prompt article.

How does length affect the odds?

The same models are documented to lose quality as the clip gets longer, which is why every tier keeps a single generation at ten seconds or under. A ten-second clip is one long shot at it. Two five-second clips are two short shots at it for the same 980,000 tokens, plus a cut point. If a camera move doesn't have to run unbroken, the short clip is safer.

Where the fixes stop

Here's the limit: hands, physics, and small type are engine limits, and no prompt removes them. The same source models also note that a prompt can be flat-out misread, so some share of reshoots has no cause on this list. There's no seed either, so a good shot can't be re-run with one word changed and the rest held fixed. The Advanced panel shows which engine each tier goes to, and it's read-only.

A practical budget is two draft attempts plus one Cinema shot — 1,680,000 tokens, still less than a single 1,872,500 Cinema Pro shot. The cheap tier absorbs the uncertainty; the expensive tier records the answer.

FAQ

Can a bad clip be fixed instead of regenerated? Only on the timeline. Cinema Studio crops, sequences, overlays text, and balances the music. It doesn't repaint frames, so a wrong face or a broken sign means a new generation.

Does the draft tier miss more often than Cinema? Tiers run different engines, so a prompt that misses on Draft can land on Cinema, and the reverse. The cheapest way to learn this is a draft image first, then a single clip at the tier the shot will ship on.

Can I see the prompt the studio actually used? Yes. In Flow Studio every scene card has a Prompt field carrying the sentence the studio wrote from your synopsis, editable before the scene is generated. The Library also keeps a Copy prompt button on every finished item.