A hundred generations for twenty seconds of film
By Syntax Motion, Vienna. Updated: 16 October 2026
AI video production is usually written about as if it were the push of a button. This post is the opposite: a workshop report with numbers from our own projects. What really happens between the description of a scene and the shot that makes it into the film.
The number nobody talks about
For every shot that makes it into the finished film, at least five are created here that nobody ever sees. For a film of one minute that adds up to around a hundred generations, of which about twenty end up in the result. This waste is not a flaw in the method, it is the method.
That sounds wasteful and is the opposite. Every attempt costs computing time instead of a shoot day, and discarding is the part that makes the quality. Whoever takes the first usable shot gets a film that looks like the first usable shot.
Why the other four fail
The reasons are the same five across projects. In this order we meet them most often:
| Reason | What is behind it |
|---|---|
| The movement is wrong | the camera moves too fast, too slow or in the wrong direction |
| Errors in the detail | hands, type on packaging, reflections, see inpainting |
| Does not match its neighbour | light or colour differ from the shot before |
| The character changes | the core problem, see character consistency |
| Correct but boring | technically flawless and still not an image anyone watches |
The last row is the interesting one. A model reliably delivers correct images today. What it does not deliver is the decision which of them tells something.
Why the prompt is overrated
A good prompt raises the hit rate, perhaps from one in eight to one in five. That is worth a lot and changes nothing about the principle. The work shifts from writing to choosing, and choosing is direction.
In practice that means we almost never work from plain text. The approved styleframe becomes the starting image, and the movement grows out of it. Half of the uncertainty is gone before the first second of film exists.
What this means for your quote
Three things follow from it, and they stand in every quote we write:
- The price depends on the number of shots, not on the length of the film. Twenty seconds with twelve cuts are more work than sixty seconds with four.
- Closeness costs. A character in close-up across five shots is a multiple of the effort compared with wide shots and details.
- The schedule has no shoot day, but a long selection phase. The date on which you see the first complete film is later than expected, and the changes before it are cheaper.
Frequently asked questions
How many attempts does an AI shot really take?
For one shot that makes it into the film, at least five are created here that get discarded. That adds up to around a hundred generations per minute of film, of which about twenty end up in the result. The ratio is remarkably stable across projects.
Why does a short ad sometimes cost more than a longer film?
Because the price depends on the number of shots, not on the running time. Twenty seconds with twelve cuts need more generations, more selection and more matching work than sixty seconds with four calm shots.
Do we get to see the discarded attempts?
On request yes, but they are rarely helpful. More useful are the motion tests that we present for approval anyway. There you see two or three versions of the same shot and we choose together.
Will this get better with better models?
The hit rate rises, the principle stays. Even at a rate of one in three the actual work would be the same: recognising which result holds up next to the others. That is direction, and no model takes it over.
Read on: All posts • How a project runs • Pricing • Glossary
How many shots does your film need?
We settle that in the first call, before a number stands in the quote.
Book a first call