How many times should I regenerate this? It’s the most honest question in AI video — and the one with the worst answers floating around. The forums say until it looks right, which is a recipe for an empty credit balance. The productivity crowd says stop at three, which sounds like a fortune cookie.
Neither answer survives contact with a real project, where one shot is the face your whole short hangs on and the next is two seconds of rain on a window. In my pipeline the answer is decided before generation starts, per shot, based on what that shot is worth and what its failures are telling us.
This article is that system: how to set a reroll budget, how to read whether another roll is even the right tool, and how to walk away from a shot before it eats the project.
Every Shot Has A Reroll Budget
Start with the premise that changes everything downstream: rerolls are a budget line, not a mood. On a live-action set, nobody shoots forty takes of an insert and two takes of the climax. Time is allocated by importance. Generated production deserves the same respect — because credits are simply time wearing a different uniform.
In my workflow the allocation happens at the storyboard, in the same pass that decides shot size and negatives. A panel gets marked with its job — hero, story, connective, texture — and its job carries a number of permitted generations. Nobody negotiates with that number mid-session at 1 a.m. That’s rather the point of writing it down while calm.
What does a budget actually buy? Sampling. Within a sound prompt and an approved frame, different rolls give you different performances of the same shot — and choosing among performances is legitimate directing. A budget buys you a fair sample of those performances: enough to pick a take with confidence rather than grabbing the first survivable one. What a budget doesn’t buy is a different shot. If the plan is broken, the budget is just the price of finding out slowly. Which brings us to the diagnostic that governs everything below.
I keep the numbers deliberately small, and the table later in this article puts my working ranges on the page. But the exact numbers matter less than the existence of a limit decided in advance. An unlimited reroll loop feels generous and behaves like a leak. A written budget feels strict and behaves like a producer.
My footage got better when I hired the producer.
Judging Variance: Are The Failures Siblings Or Strangers?
The single most useful reroll skill is reading variance — comparing your takes to each other and deciding what their pattern means. I lay attempts side by side in Premiere, watch once at speed, then step through the failures. Two patterns emerge. And they dictate opposite actions.
Sibling failures repeat. The hand breaks at the same contact point in three takes. The face drifts on the same head turn. The background warps along the same stretch of the camera move. Siblings are the prompt — or the frame, or the shot design — producing the failure reliably. Another roll will produce it again with minor costume changes. When I see siblings, the reroll budget is suspended, not spent: the next move is a fix, aimed at the cheapest upstream cause. Simplify the action. Shorten the move. Tighten the framing. Rebuild the starting frame.
Rolling against siblings is the most common credit fire I see in other creators’ sessions. And it burns because each new take looks just different enough to feel like progress.
Stranger failures, by contrast, arrive fresh each time. One take flickers, the next runs a touch soft, a third lands the gesture a beat late — each version mostly right, flawed in a new place. Strangers mean the plan is sound and you’re sampling performances. That’s precisely what seeds and rerolls exist for. Spend the budget. Choose the best execution. This is also where taste legitimately enters: among three good-but-different takes, pick the one with the performance you want. That’s directing, not gambling.
The diagnostic discipline is simply refusing to confuse the two situations. If you take one tool from this article, take the sibling test. It runs in under a minute and it has saved me more credits than every other habit combined — closely followed by the three-roll discipline I describe in the seed-versus-prompt article in this series.
The Signal That The Prompt, Not Luck, Is Broken
Some signals are loud enough to call from across the room. If the failure involves something you never described — a person who was never cast, an object never specified, weather you never mentioned — the model is improvising into a vacuum, and improvisation is fixed with words, not rolls. If the described action itself never quite happens, take after take — the reach that never reaches, the turn that never completes — the action as worded exceeds what the shot can perform. Simplify it. If the failure lives in the first second of every take, suspect the starting frame before anything else, because generations inherit their opening like a first impression.
And if the failures are getting more varied as you roll rather than less? That’s not luck improving. That’s an underspecified prompt giving the model room to wander — and wandering widens with sampling.
The inverse signal matters too, because over-fixing is its own tax. If each take is closer than the last — defects shrinking, performance settling — you’re watching a prompt converge under sampling, and a couple more in-budget rolls are honest spending. The pattern to trust is direction: converging takes justify rolls; identical failures forbid them; diverging failures indict the words. I write the diagnosis in the project notes in plain language — “siblings, hand contact, simplify action” — because three shots later I won’t remember which take taught me what, and the note is the difference between a pipeline and a pile of renders. This is the same evidence habit behind my seed methodology: record the decision at the moment it’s made, or pay to remake it.
Shot Value: Hero Shots Earn More Rolls
Not all seconds of finished video cost the same. So they shouldn’t be sampled the same. A hero shot — the close-up the story turns on, the product reveal, the frame that will be screenshotted — earns a deeper budget, because the audience will live in it. Connective shots — walking from door to table, the glance that bridges two scenes — earn a shallower one, because the audience passes through them and the edit can help them pass gracefully. Texture and atmosphere shots earn the shallowest sampling of all. Not because they matter least, but because their success criteria are the loosest and three takes almost always contain a usable one.
There’s a second axis creators forget: replaceability. Some shots, if they fail, can be covered — a cutaway exists, the beat can move to the neighbouring shot, the scene survives. Others are load-bearing. Nothing else can carry that story beat. Load-bearing hero shots are where I allow the budget to flex upward after a genuine prompt fix resets the sampling. Irreplaceable and important is the only combination that earns real persistence. Everything else gets its budget and a graceful exit.
And a confession to keep this honest: shot value must be judged at the board, not during the session. Mid-session, every shot feels like a hero shot — because you’re staring at it. The board, made calmly, knows which panels carry the piece. Trust the earlier, cooler judgement. That’s half of what the storyboard-first workflow is for: making production decisions at the moment you can afford to make them well.
A Reroll Budget Table By Shot Type
Here are the working ranges from my own pipeline. They’re starting postures, not commandments — your tools, costs, and standards will bend them — but they’re calibrated to real projects rather than forum bravado, and they assume an approved frame and a finished prompt before roll one. Rolls after a genuine prompt or frame fix start a fresh count, because a new plan deserves a new sample.
| Shot Type | Planned Rolls | Stop And Fix When |
|---|---|---|
| Hero close-up / key reaction | 4–6 | Same drift or defect in 2 takes running |
| Dialogue or performance medium | 3–5 | Action not landing, or face drifting identically |
| Product hero shot | 3–5 | Shape or label failing the same way twice |
| Walking / movement connective | 2–3 | Feet, gait, or background failing as siblings |
| Wide establishing shot | 2–3 | Geometry warping along the same move |
| Cutaway / insert / texture | 1–2 | Almost never — take the best and move on |
| Atmosphere / weather / background plate | 1–2 | Rarely; loose criteria, generous supply |
Read the right-hand column as seriously as the numbers. The budget is a ceiling; the sibling rule is a tripwire that can end sampling early — and it outranks the ceiling every time. Six permitted rolls don’t entitle a broken shot to six generations. Two identical failures cashier the plan regardless of what the table allows.
One more line of the system: leftover budget doesn’t transfer. Finishing a cutaway in one roll doesn’t grant the hero shot a seventh. Transfer feels efficient and behaves like scope creep; the totals quietly inflate project-wide. Unused budget is profit. Bank it and finish earlier.
Walking Away: Cutting The Shot From The Board
The last discipline is the one nobody wants to hear: some shots should die. If a shot has consumed a fix, a fresh sample, and still produces sibling failures, the evidence is in — as designed, it exceeds what this generation can carry, and further rolls are a donation. In my pipeline that verdict triggers reboarding, not more spending: split the beat across a cut, move the action to a closer and simpler shot, hand the story point to a neighbouring panel, or cut the beat entirely and let the edit imply it.
Audiences are brilliant at implying. They can’t unsee a melted face — but they’ll happily never know a walking shot once existed.
Walking away is easier when the decision is procedural rather than emotional. Which is why the budgets and tripwires exist above. The rule fires; the shot gets reboarded; nobody argues with the footage at midnight. I also log the cut and the reason. Those logs have quietly redesigned how I write boards — I now design simpler walks, shorter takes, and fewer three-action shots from the start, because the log showed me exactly which designs kept dying. The QC checklist catches defects; the reroll log catches expensive ideas, one stage earlier.
So — how many seeds should you try before rewriting the prompt? As many as the shot earned at the board, and not one more than the sibling rule allows. Give sound plans a fair sample, give broken plans a fast fix, and give impossible shots a dignified cut. Your credits will start behaving like a production budget instead of a slot machine.
And strangely, the footage gets better too — because every generation you do run was planned by someone who intended to stop.