Re-roll, never repair: the rule that governs an AI production budget

01Should I regenerate a bad AI clip, or edit it?

Regenerate it. Editing a generated take treats the symptom you can see and leaves the cause, because the cause sat in the instruction rather than in the pixels. The exception is narrow and worth knowing: anything added after the take - a crop, a trim, a grade, a caption, a sound - is a genuine edit, and re-rolling for one of those wastes money.

The confusion lives in the word wrong. A take can be wrong in two unrelated ways. It can be wrong in the frame, which is a rendering fault, and those are usually loud and rare. Or it can be wrong in what it was asked to do: the performer emotes at nothing, the camera drifts where it was told to hold, the light arrives from a window the room does not have. Nothing downstream un-decides any of that.

The row people argue with is the first one, and they are right to. A re-roll costs real money and a crop costs none. Which is exactly why the re-roll has to be a decision with a name on it rather than a habit, and why the arithmetic of the whole set matters more than the price of one attempt. That sum is in what failed renders actually cost, and the route-level version of the same question sits in what AI ads, UGC creators and an agency each cost.

02Why repairing looks cheaper than it is

Because the first repair is cheap. It is always cheap. You are twenty seconds into a crop that hides the hand, and it feels like you just saved a generation. The bill turns up later, in three places nobody traces back to the crop.

The first is that a repair has no definition of done. There is a documented name for what happens next: escalation of commitment, described by Barry Staw in 1976, where people keep investing in a course of action as the evidence against it piles up, precisely because they have already invested. It reads like organizational theory and it describes a Thursday afternoon in an edit suite exactly.

The second cost is that the person best placed to repair a take is the worst placed to judge the repair. We measured this rather than assumed it: on the same films, a builder's own scores ran between 36 and 76 points above the owner's. Self-scoring measures compliance with the brief, which is the one thing the builder can see, and taste is the thing it cannot.

The third cost is the one we paid for and wrote down as a rule: a lint is a floor, not a target, so when a check fails you fix the thing rather than the number. The soundtrack we tuned green and shipped anyway, and what the listener said about it, is in sound is 40% of the ad.

03How do you tell a fixable note from a re-roll note?

Ask what the note names. If it names something added after the generation - a crop, a trim, a grade, a caption, a sound - it is fixable, and you should fix it. If it names what the performer did, what the camera did, what the light did or how a material behaved, it is a re-roll, because all of those were decided inside the take.

Two tests, ten seconds each. Could this note have been written before the take existed? If yes, it is a direction note and it belongs in a brief, not in an edit. And second: if you applied the fix, would the thing being complained about still be true? A crop that hides a bad hand still has a bad hand under it, and the hand reappears the moment somebody asks for the wider version.

None of this needs you to be a video person. It needs you to describe what you saw, in order, in your own words, and to resist proposing the fix in the same breath. If that part is the part you find hard, we wrote the longer version in how to review AI creative when you are not a video person.

04Every re-roll is a decision somebody has to authorize

In our pipeline a bad take cannot be re-rolled by whoever noticed it was bad. The re-roll is a request: one line on what failed, one line on which single variable changes, and the price of the generation. Somebody holding the budget says yes or no, and nothing runs before they do.

This sounds like paperwork and does the opposite of what paperwork does. It exists because generation got cheap enough to become a reflex, and a reflex has no upper bound. Between any two human checkpoints our spend is capped and our compute is capped at one cheap mockup pack per creative. The promise is not that we will never be wrong. The promise is that wrong is cheap.

The side effect is the part we did not predict. A good share of re-roll requests die while they are being written, because naming the single variable you would change forces you to notice that you do not have one. A request nobody can finish writing was never a re-roll. It was a feeling about a take.

05How do you write a re-roll brief that changes one thing?

Copy the brief that produced the failed take, change one line, and mark which line you changed. That is the method. A re-roll brief that rewrites everything is a new brief, and when the result comes back better you will not know which change did it, so the next take is another guess at full price.

Two things belong in it that most people leave out. The failure sentence, in a viewer's words rather than a maker's. And the pace: say what happens and how fast, never a bare duration. Asking for seven seconds of a two-second action is an instruction to move slowly, which is how you get a re-roll that is worse than the take it replaced.

RE-ROLL 04b   (replaces take 04 - same beat, same slot, same length)

FAILED   She smiles at nothing. No cause for the expression exists
         in the frame, so it reads as performance rather than reaction.

CHANGE   One variable: the cause. She reacts to the pouch opening in
         her own hands, in frame, at 0:01.

HOLD     Everything else. Same room, same window light from camera
         left, same wardrobe, same lens, same grade, same length.

PACE     The open takes about a second. Her face changes after the
         sound, not before it. Neutral until then - neutral is a
         legitimate state and usually the correct one.

COST     One generation.        Authorized by: ______

That example is a real class of failure rather than an invented one: four takes from one shoot died together for performed expression with no cause in the frame, and the account of it, with the rule that came out of it, is in keeping one performer continuous across six shots. If you have a reference you like, hand it over rather than describing it: the reference video is the specification, and the pace question has its own piece in duration is the action's pace, not a length.

06The loop where every round costs money and nothing improves

We once shipped a fix round that verified the wrong thing entirely. The note said a label was pointing at the wrong object. Our fix confirmed that the leader line met the text, which it did, and never once asked what the arrow was pointing at. The reply came back in four words: you didn't fix anything.

That is the mechanism of the loop. Each round validates itself against its own account of the complaint rather than against the complaint. Everyone is working, money is moving, and the thing the client actually said has quietly dropped out of the process. Five moves break it, and not one of them needs a better model.

Repair is the only creative decision that gets more expensive the more of it you buy. Every other line in a production budget has a floor and a ceiling. This one has a floor of nearly zero, which is why people start, and no ceiling at all, which is why they cannot stop.