It would take 30-60 seconds for one image. But the chance of getting the right one is slim. You’d probably generate hundreds, settle on one, then modify that one further. Depending on how precise you want it, it could take over an hour of GPU use. Services that do it online are doing the same thing in the background, maybe generating even more and iterating and evaluating them. All while using many GPUs at once. This isn’t so bad for an image, but videos are way more resource intensive.