The Retry Problem

Published October 2026 · 11 min read

Ask someone what AI video costs and they will quote you a per-second rate. That rate is real, but it describes a world where every generation is usable. In the world you actually work in, you pay for every attempt and keep perhaps one in three.

That gap — between the price of a generation and the price of a usable generation — is where most of the budget goes. Not the model choice, not the resolution, not the subscription tier. Retries.

This article is about reducing them systematically rather than hoping each run goes better.

What a retry actually costs

Take a simple case: a 5-second shot at 1080p on Veo 3.1 Lite, at $0.10 per second.

The cost of the shot you use is the same in all three cases. What changed is how much you paid to find it. A project with a six-attempt average costs three times a project with a two-attempt average — for identical output. This is the single largest controllable variable in AI video production, and it is almost never the thing people optimise.

Multiply across a 30-second piece of six shots and the difference between a 2× and a 5× retry rate is the difference between roughly $6 and $15 on the same model. Across a year of weekly publishing, that is thousands of dollars.

Six causes of retries, and which ones you can fix

Not all retries are equal. Sorting them by cause is the first step to reducing them, because only some are your fault.

1. Prompt gaps (fixable)

The model invented something you did not specify, and it was wrong. Skin texture, light direction, hand position, hair movement. Every empty slot in the prompt is a coin flip, and enough coin flips guarantee a retry.

This is the largest fixable category. The six-slot framework exists specifically to eliminate it.

2. Contradictory instructions (fixable)

Two clauses that cannot both be true — “hard midday sun” and “soft diffused shadows”, “static tripod shot” and “handheld energy”. The model averages them into something that satisfies neither. These are easy to miss because each clause reads fine on its own.

3. Asking for too much motion (partly fixable)

Complex action has an inherent failure rate. You can reduce it by shortening the shot or splitting the action, but you cannot make a five-second fight scene reliable. Accept the cost and budget for it.

4. Model randomness (not fixable)

Even a perfect prompt produces different results on different runs. Some of that variation is genuinely useful — it is how you find the best take. The trick is not to eliminate it but to stop re-rolling for marginal gains.

5. Wrong model for the shot (fixable by routing)

A shot with dialogue on a silent model, or fast action on a model that handles motion conservatively, will fail repeatedly for reasons no prompt edit can fix. Switching models often resolves in one attempt what five prompt rewrites could not.

6. Unclear success criteria (fixable, and the sneakiest)

You keep generating because you have not decided what “good enough” looks like. This produces endless takes that are all acceptable. Decide before you generate: what must be true for this shot to be usable? Then stop at the first one that meets it.

Realistic retry rates by shot type

From our own project logs, averaged across models. Treat these as planning baselines, not guarantees.

Shot typeTypical attemptsNotes
Wide landscape, no subject1–2Forgiving; almost anything reads as valid
Product close-up, static1–2Low motion, low failure surface
Single person, subtle motion2–3Face and hands are the risk
Talking head with dialogue2–4Lip-sync and eye direction
Two people interacting3–5Both must stay coherent
Fast action / sport4–6Motion coherence degrades quickly
Hands doing detailed work4–7Highest failure surface in the medium
Crowd scene5–8Background figures multiply the error

The pattern is clear: failure rate scales with the number of things that must stay coherent simultaneously. A landscape has one thing. A crowd scene has thirty.

The three-attempt rule

This is the most useful habit we have adopted. If a shot fails three times in a row, stop generating and diagnose. Do not run a fourth.

The reasoning: three independent failures is strong evidence of a systematic problem, not bad luck. Either the prompt has a gap, the model is wrong for the shot, or you are asking for something the medium cannot do. A fourth attempt has the same expected outcome as the first three.

What to do instead, in order:

  1. Re-read the prompt for gaps. Is there an unspecified dimension the model is guessing at?
  2. Check for contradiction. Do any two clauses pull in different directions?
  3. Reduce the ask. Cut the motion, pull the camera back, shorten the duration.
  4. Switch models. Same prompt, different engine. This resolves a surprising number of cases.
  5. Change the approach. If the shot cannot be generated, can it be composed from two simpler shots, or covered by a different angle?

In practice, step one or two resolves most three-failure cases. The expensive mistake is skipping straight to attempt four, five and six, which is what most people do.

Five system-level changes that cut retries

Write the prompt in slots, not sentences. A blank slot is visible in a list and invisible in a paragraph. Fill all six before generating — this alone removes the largest fixable cause.

Generate at draft resolution. Compose and iterate at 720p, then re-render only the keeper at 1080p. You pay half price for every failed attempt, which halves the cost of the retry tax without reducing it.

Batch by shot type. Doing all close-ups in one session keeps you in one mental model of the prompt and one set of vocabulary. Switching between shot types repeatedly produces more gaps.

Keep a failure log. One line per abandoned shot: what you asked for, what came back, what fixed it. After twenty entries you will see your own personal failure pattern, and it is usually two or three recurring issues.

Define “good enough” before generating. Write it down. “Face in focus, hands not visible, light from the left.” Stop at the first take that meets the criteria. This is the single cheapest change on the list.

When to abandon a shot entirely

Some shots are not worth the credits. Abandon when:

An abandoned shot is not a failure. It is a decision that the budget is better spent elsewhere, and it is usually correct.

What a retry log actually looks like

A log is only useful if it is fast to write. Ours is three columns, one line per abandoned attempt, and it takes about ten seconds per entry.

shot | what came back wrong | what fixed it

A real stretch from last month, anonymised:

After twenty lines, patterns appear. Ours were: hands, unspecified light direction, and over-large motion. Three recurring causes accounted for the majority of wasted generations — and all three were fixable with prompt changes, not with better models.

The log also produces something less obvious: a personal list of shots you should stop attempting. The “child running” line is a decision, not a failure. It says that this particular shot type is not worth the credits in this workflow, and it prevents you from re-learning that lesson every few weeks.

Keep the log wherever you already work — a notes file, a spreadsheet, a text file next to the project. The format matters far less than the habit of writing one line instead of silently re-rolling.

The arithmetic worth remembering

Two projects, same model, same duration, same resolution. One averages two attempts per shot, the other averages five. The second costs 2.5× as much for output of identical quality. The only difference is process.

This is why the cheapest model is not always the cheapest project, and why a $0.40/s model used with discipline can cost less than a $0.05/s model used carelessly. The cost breakdown shows how the multipliers stack; this article is about the one you control most directly.

Related reading: the nine tells that make AI video look fake are, from a budget perspective, nine causes of wasted generations. The 12 costly mistakes page covers the wider set of budget leaks, and the prompt length checker flags the dilution zone where added words start causing rather than preventing retries.