I generated one five minute clip last week and it burned more credits than every short ad I had made in the days before it. The clip was not even usable. I assumed I had been billed for quality, or for resolution, or for some queue priority I had accidentally opted into. It was none of those.
Lip-sync compute scales with duration, not with whether the result was any good. A single long take has to re-solve mouth shapes across every frame it contains, so a five minute take is not five times the work of a one minute take in the way a longer render usually is. And if the take comes back unusable, which long takes often do, you paid for all of it before you found out.
The fix is boring and it works: generate in five to eight second beats and stitch them. Sync comes out cleaner because each beat is solved over a short window, and a bad beat costs you one beat instead of the whole ad. This is also how the short ads had stayed cheap without me noticing why.
The other half of it is picking the model for the job rather than for the spec sheet. CastTake's model and pricing page lists every model it can run with the rate per second, and the spread is wider than I expected: Grok Imagine 1.5 at 480p is 4 credits a second, MiniMax H3 is 14 at 720p and 22.4 at 1080p, one credit being one US cent. A five second clip is 20 credits, 70, or 112 depending purely on that choice.
Stitching has a cost I should name, because it is the part nobody mentions when they recommend it. You own continuity yourself now, and keeping a presenter recognisable from one beat to the next needs a reference image rather than luck.
So the order matters more than the choice. Generate ten hooks on the cheap model to find the one that works, since 480p is already past the resolution most of a phone feed is watched at, then finish the single winner on the expensive one. Running the whole batch at 1080p costs five and a half times as much and does not make a weak script work.
Top comments (0)