Generation controls

    Generation time

    Also called Latency, Wait time.

    Generation time is how long a job takes to finish once you submit it — driven by resolution, step count and model size — and it is routinely confused with duration, which is how long the finished clip plays.

    The two measure different things entirely. Duration is a property of the output: it is still true a week later, and it is the number you picked from the model's menu. Generation time is a property of the process that made it: it exists only once, while the job runs, and nothing about it survives into the finished file. A four-second clip and a sixty-second clip can take roughly the same wall-clock time to render, and a short clip on a heavy model can take longer than a long one on a fast variant.

    What actually drives the wait is the same handful of factors the parameter and architecture entries already name individually. More pixels mean more compute per step, which is resolution's cost. More sampling steps mean more refinement passes, which is close to a linear multiplier on time by that entry's own account. A distilled or fast model variant is quicker specifically because it was trained to converge in far fewer of those passes — the saving is architectural, not a shortcut through a queue.

    None of that is about queueing or concurrency, which is a separate, account-level question this entry does not attempt to answer. Running several jobs at once does not make any single one of them faster; it only means they tend to land around the same time as each other.

    In practice

    • A short clip is not automatically a fast one; resolution and step count matter more than the length of the output.
    • A fast or distilled model variant is quicker because it converges in fewer passes, not because anything about the job itself changed.
    • Running several jobs in a batch does not speed up any single one of them; it just clusters when they each finish.

    The mistake to avoid

    Reading a slow generation as a broken job. A heavy resolution or a full-size, non-distilled model at a high step count simply takes longer on the same footage than a lighter setting would — that is the trade the setting made, not a fault.

    Related terms

    The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.