Why a Generation Takes Time

Your own limit is a number of slots, not a speed

Your account runs a set number of generations at the same time. Start more than that and the extra ones park: they sit in a waiting state, and as soon as one of your running jobs finishes, the next one is promoted into the free slot automatically. You do not have to come back and press anything.

Parking is not a failure. Nothing errors, nothing is lost, and the credits for a parked job are already held aside — frozen, not spent. If you cancel it before it starts, they come straight back.

How many slots you get depends on your plan. A paid plan raises the number of jobs that can run at the same time, and the top plans also put your jobs in a priority queue ahead of everyone else's. What no plan changes is how long a single generation takes — that part is the model's, and it is the same model either way.

Three short interactive steps are deliberately never parked: writing a script with AI, suggesting a style, and matching a character voice. They run even when your slots are full — they are seconds-long text calls, and it would be absurd for them to queue behind your own ten-minute video. They do still occupy a slot while they run, so starting one when you are already at your limit can briefly park the next thing you submit.

The shared limit

Behind your slots there is a second cap: how many jobs the whole site may have in flight at a model service at once. That counter is shared across every machine we run, so it is genuinely global — when the site is busy, your job waits behind other people's.

This is why the same segment can start instantly at one hour and wait at another, with nothing about your project having changed.

Where the number on screen comes from

The estimate is arithmetic, not a forecast. We take how many jobs are queued on the same model service at your priority or above, divide by how many the workers there can run in parallel, and multiply by how long the last fifty successful jobs on that service took. Then we pad the result by a fixed safety factor, because an estimate that runs out before the job does is the worst kind of wrong.

So it is backward-looking by construction. It gets less accurate exactly when things are unusual — which is also when you are most likely to be watching it.

What actually takes the time

Almost all of it is the model. Our own part — assembling the prompt, resolving which reference images to attach, writing the result back — is a couple of seconds at most.

For a rough sense of scale, an internal measurement taken in August 2026 put the average at around a minute and a half for a setup image, around nine and a half minutes for one segment of video, and around thirteen minutes to composite a full episode. Treat those as orientation, not as a service level: they were one measurement, they move with the tier you pick and with load, and nothing keeps them current.

The one number under your control is the tier. A cheaper, lower-resolution tier is usually also a faster one, which is why the fastest way to find out whether a shot works is to try it cheap first and only then spend the good tier on it.

If it seems stuck

A job that has genuinely gone wrong does not sit silently — it ends. If the model never received the request, or answered that it was busy or rate-limited, the credits are released back to you. If the call timed out or the connection broke mid-flight, we cannot tell whether the model ran and billed us, so the credits stay frozen rather than being either charged or refunded, and a person settles it by hand.

What that means for you: a balance that looks short after something failed is usually held, not gone. The credits panel in the top bar has the ledger, and every freeze, charge and release is a line in it.

FAQ

Updated September 21, 2026