The most expensive button in AI video is “generate again” pressed without a thought in between. I know. Because I used to press it professionally. A shot would come back wrong, I’d change three words at random, add “high quality” like a prayer, re-roll, get a differently wrong version of the same mistake, and call the session “iteration.”
It wasn’t iteration. It was slot-machine play with a vocabulary. And my credit balance could tell the difference even when I pretended not to.
What replaced that habit is a diagnostic method: read the failure, name what the model actually did, map it to a cause, change one variable, re-test. Two disciplined passes beat ten random ones — in credits, in time, and in what you learn. This guide lays the method out the way I run it in my own Seedance and Veo pipeline, where a failed generation is treated first as information about the board, then about the words, and only lastly about the settings.
Stop Re-Rolling And Start Reading
The urge to re-roll instantly is understandable. Generation feels cheap per click. And sometimes the next seed really does fix it.
That “sometimes” is the trap. A re-roll without a change tests nothing except variance — whether the same instruction happens to land better on a different random path. When it lands better, you learn nothing you can reuse. When it lands the same, you’ve paid twice for one data point.
Reading a failure properly takes about ninety seconds. And it starts with a rule that sounds harsh: the model probably did something reasonable with what it was given. Not always — models artefact, faces collapse, physics fails. But a large share of disappointing generations are faithful executions of ambiguous, overloaded, or contradictory prompts, and treating them as random malfunctions blinds you to instruction errors you’ll repeat on the next shot, and the next project.
There’s a pipeline argument too. In a storyboard-first workflow, a bad generation may be reporting a bad board: a frame with two subjects competing, a contact nobody drew clearly, an action with no visible end state. Re-rolling the prompt over a broken board polishes the wrong stage. The diagnostic starts upstream on purpose — frame first, then words, then settings — because that’s the order the mistakes were made in and the order of cheapness in fixing them.
So the working posture is this: every failure is a sentence the model wrote back to you about your prompt. Your job is to read it before replying. The rest of this article is how to read it, how to classify what it says, and how to answer it with a change that teaches you something.
Read The Failure: What Did The Model Actually Do?
Watch the failed clip three times, each time with one question. Full speed first, for the overall betrayal: what did you ask for, in one sentence, and what happened instead, in one sentence. Write both sentences down. The gap between them is the diagnosis’s raw material — and writing it prevents the vague “it’s just off” feeling that leads straight to random tweaks.
Second pass, slow or scrubbed: where exactly does it depart from the intention? First frame already wrong, or correct until second three? A failure that exists at frame one is a description or reference problem — the model started in the wrong place. A failure that develops is a scheduling or load problem: it started right and couldn’t sustain what it was carrying. That single observation halves the suspect list before you’ve changed a word.
Third pass, muted, only the corners of the frame. Backgrounds, hands, edges, the subject’s far side. Generations often perform the centre honestly and improvise the periphery. If the failure lives in the periphery — breathing walls, a second figure inventing tasks, reflections misbehaving — the main action may be blameless, and the rewrite target is background control, not the hero description.
Then complete the reading with an unfashionable question: which part worked? A failed clip usually contains a working component — the light behaved, the camera move tracked, the first gesture was clean. Naming the working part matters because random rewrites destroy working parts along with broken ones. A diagnostic rewrite protects what worked by leaving its wording untouched, so the next generation can confirm you fixed the failure rather than traded it.
I keep these readings brutally short. Three lines at most, attached to the generation in the project folder. If reading a failure takes longer than generating it did, the method will be abandoned by Friday. Ninety seconds, three lines, one named suspect. That’s the whole liturgy.
The Five Failure Families And Their Prompt Causes
Across my own failure logs, nearly everything sorts into five families. Each family has a characteristic look, a usual prompt cause, and a first fix that isn’t “try again.”
Family 1: Drift — The Subject Slowly Becomes Someone Or Something Else
Faces soften and re-form, wardrobe items redesign themselves, a product’s proportions wander. Drift is a description-load problem: the identity was described once and then asked to survive motion, camera change, and duration without reinforcement — or the description contained soft, contestable words the model re-interprets continuously. First fix: strengthen and stabilise the identity wording (same words, early position, concrete features) and shorten the distance the generation asks the identity to travel. New camera, same person, shorter move. Drift across a sequence, rather than inside one clip, is a consistency-system problem, handled at the reference level as in the character consistency guide.
Family 2: Melt — Contact And Complexity Come Apart
Hands merge with props, two actions blend into a gesture soup, a bite dissolves. Melt is an overload report: too many simultaneous systems, contact without named touch points, speed beyond what the contact can survive. First fix is subtraction, not addition. Fewer actions, slower verb, simpler grip, shorter duration. If a rewrite for melt adds words, they should be words that assign and restrain — which hand, which handle, how slowly — never words that decorate.
Family 3: Disobedience — The Model Plainly Does Something Else
You asked for a leftward pan and got a push-in. The subject exits the wrong side. The second speaker delivers the line. Disobedience is often a contradiction report: two instructions that can’t both win, with the model electing the one phrased more strongly, placed earlier, or repeated. Read your prompt for internal elections. Camera words versus subject-motion words are the usual candidates — a subject “rushing toward camera” will outvote a timid “slow dolly.” First fix: remove the losing instruction’s competitor, restate the winner once, plainly. Adding emphasis adverbs to a contradicted instruction is arguing louder in a language the model settles by vote count.
Family 4: Blandness — Everything Works, Nothing Is Alive
The technically clean, emotionally empty clip. Nothing melted, nothing drifted — and nobody would watch it twice. Blandness is a specificity deficit: generic verbs, no pace, no endpoint, light without a source, style by compliment. First fix: sharpen one layer only — usually the verb, given a destination and a pace word — and name the light source. Blandness rewrites should add decisions, not adjectives. A shot can’t be rescued by being told to be more beautiful. It can be rescued by being told what happens in it, more exactly.
Family 5: Artefact — The Frame Itself Breaks
Flicker, warping geometry, text dissolving, texture boiling. Artefacts split into two sub-causes worth separating. Load artefacts ride along with melts — too much demanded of the pixels — and yield to the same subtraction. Settings and source artefacts persist even on simple prompts, and point at the input rather than the prose: a low-quality or busy first frame, an over-demanding style anchor, a duration setting stretching a short motion too far. First fix for persistent artefacts on a simple prompt is to test the same prompt on a cleaner frame or plainer look before rewriting words that may be innocent. The quality-check method is useful here for documenting exactly which frames break, so artefact fixes can be verified instead of vibes-based.
Most failed clips belong primarily to one family with a secondary guest. Name the primary. Fix the primary. A rewrite that addresses three families at once is just a new prompt wearing the old one’s name — and its result will teach you nothing about any of them.
Change One Variable Like You Mean It
The rule sounds simple and is violated constantly. Including by me on tired evenings. Change one variable per pass.
One variable means one layer of the prompt, adjusted decisively — the verb and its pace, the camera line, the identity wording, the duration setting — not one word fiddled at the margins.
“Decisively” matters. Changing “walks” to “walks slowly” and re-rolling tests almost nothing, because the variation between seeds is larger than the variation you introduced. A diagnostic change should be big enough that its effect, if any, is visible above seed noise: cut a whole action, swap the camera move for a locked frame, halve the duration, replace the verb family entirely, or remove the style anchor that may be taxing the render. If the result improves, you’ve found a load-bearing wall. If it doesn’t, you’ve eliminated a suspect — which is also progress. Recorded progress, not spun wheels.
Keep everything else frozen, including the seed where the tool allows it. A frozen seed with a changed prompt is the closest thing this craft has to a laboratory control, and the seed methodology on this site exists precisely to make such comparisons honest. Changing the prompt and the seed together produces a new sample from a new distribution — possibly prettier, wholly uninformative. Pretty and uninformative is the most expensive combination in generation, because it encourages you to keep a prompt whose virtues you can’t explain or repeat.
There’s a credit discipline attached: two controlled passes, then reassess the board. If two decisive, single-variable rewrites haven’t moved the failure, the problem is unlikely to be phrasing. It’s the frame, the shot design, or the ask itself. Regenerating the still, splitting the shot, or reframing the beat aren’t admissions of defeat at that point. They’re the diagnosis concluding. The expensive failure is the sixth random rewrite of a shot that wanted to be two shots since the board.
The Two-Pass Rewrite Protocol
Here’s the whole method as I run it, start to finish, on a failed generation.
Pass zero: read. Three viewings, three lines of notes. Asked-for sentence, got sentence, named family, working part identified. No typing in the prompt box yet.
Pass one: subtract or stabilise. Most failures are overload or drift, so the first rewrite usually removes or anchors rather than adds. Cut the secondary action. Name the touch points. Freeze the camera. Stabilise identity wording, copied verbatim from the project’s locked description. Make the one decisive change the family calls for, keep the seed if you can, generate once. Read the result the same way: did the named failure move? Did the working part survive?
Pass two: sharpen. With the load fixed, add the decision the shot was missing: the endpoint, the pace word, the named light source, the rest state at the end. Pass two is where blandness gets treated — because sharpening an overloaded prompt is decoration on a fire. Generate once. Read again.
Two passes, three generations total including the original failure, and you either have the shot or have evidence the shot design must change. Compare that with the re-roll habit’s typical ledger: five to eight generations, no notes, a settled-for clip whose success nobody can explain. The protocol isn’t slower. It only feels slower because the thinking happens before the spending instead of being replaced by it.
One refinement for sequences: diagnose on the cheapest representative. If four shots share a prompt skeleton and all four drift, don’t run the protocol four times. Run it on one, fix the shared wording, then regenerate the family. Prompt skeletons fail as families too — and repairing the template repairs the series. Which is why this method lives inside a template-driven prompting system rather than as per-shot improvisation.
Keeping A Failure Log That Pays Rent
A diagnostic method without notes evaporates. The failure log is what converts individual fixes into a personal craft knowledge that compounds across projects — and it only pays rent if it’s light enough to maintain on tired evenings.
Keep it at three fields per entry. The ask: the one-sentence intention, or the prompt itself pasted — pasting is faster and more honest. The family and suspect: drift, melt, disobedience, blandness, artefact, plus the single cause you judged likeliest. The fix that worked: the decisive change and its result, in a few words. “Cut the wave, locked camera, landed” is a complete and valuable entry. Screenshots optional; the failed clip already lives in the project folder if it’s named sanely.
Name files so the log can find them: shot number, take number, family letter if you like. When a new project hits a familiar wall, search the log before searching the internet. Other creators’ advice is generic by necessity. Your log is specific by construction — your model versions, your subjects, your verbs, your recurring sins. Mine tells me, with embarrassing clarity, that I overload two-person shots and under-specify endpoints when I’m rushing. Knowing that, I now check those two things before generating anything with two people in it. That pre-flight habit, bought entirely by the log, saves more than the log ever cost.
Review the log lightly at project boundaries. Promote repeated fixes into your template: if “name the steadying hand” rescued three contact shots, it stops being a fix and becomes part of how you write contact prompts on the first try. Retire entries that model updates have cured. A log that only grows becomes archaeology; a log that promotes and retires becomes a system.
The honest closing is this. Failed generations are tuition, and the only question is whether you collect the education. Random re-rolls pay the tuition and skip the class. A ninety-second reading, a family name, one decisive change, and a three-line note cost almost nothing and return a sharper prompt, a fixable shot, and eventually a personal rulebook no course can sell you — because it’s indexed to your own footage.
The model will keep writing back to you in failures. Learn to read them, and it becomes the most candid collaborator you have.