A hundred videos in one run without babysitting it

A batch is one line per video. You paste a hundred lines, they go out together, and they land as they finish. The mechanics take a minute to learn. Writing a hundred lines that produce a hundred usable clips is the actual skill, and almost all of it is decided before you press anything.

Vary four things, hold the rest

The failure mode of a large batch is not that clips fail. It is that a hundred clips come back and no two of them look like they belong to the same set, so you spend the afternoon you saved on grading and re-cutting.

The fix is to decide, before writing a single line, which parts of the sentence are variables and which are constants. In practice there are only four fields worth varying:

• The subject. What is in frame. This is the one you are actually making a hundred of.
• The setting. Where it is. Vary this second, and vary it less than the subject — five settings across a hundred clips reads as a series; a hundred settings reads as a stock library.
• The action. What the subject does. Keep the verbs in the same register: if half your lines say walks and half say erupts, the set has two tones.
• The hook, if the clips carry text or narration. One line each, written as a set, not one at a time.

Everything else — shot size, camera move, light, speed, material vocabulary — is a constant, written once and repeated verbatim on every line. That repetition is what makes a hundred separate generations read as one body of work.

Constant, on every line:
  Medium shot, static camera. Overcast daylight.
  Muted colour, matte surfaces. Slow.

Variable:
  a baker sliding a tray into an oven
  a welder lowering a visor
  a florist stripping thorns from a stem
  ...

One line each:
  Medium shot, static camera. A baker slides a tray into
  an oven. Overcast daylight, muted colour, matte
  surfaces. Slow.

Write the constant by hand first

Run the constant on its own, with three different subjects, before you generate the other ninety-seven. Three clips is three video units and about a minute, and it is the only way to find out that your lighting clause produces something greenish before it produces ninety-seven greenish things.

What happens when one fails

Failures in a batch are per generation, not per batch. A line that fails is refunded and the other ninety-nine carry on; you get the failures back as a list you can re-run without touching anything that worked.

The pricing follows the same logic. A batch of a hundred is priced as a hundred, not as one — a batch is not a discount mechanism, it is a scheduling one. Images are two credits each, so a hundred stills is two hundred credits. Video spends one unit of your video allowance per clip, and the allowance is checked against the whole batch before any of it is submitted, so you are never left with a half-charged run.

Choosing the model for a batch

Batches are where model choice stops being a preference and starts being arithmetic. Omni Flash renders 4, 6, 8 or 10 seconds; Veo 3.1 Lite renders 8 and only 8. If your hundred lines are cutaways that want to be four seconds long, that is a forty percent difference in total runtime and the choice makes itself.

If the hundred lines all contain a person, take the eight seconds and use Veo 3.1 Lite. A batch of a hundred where fifteen have a face that fell apart is not a batch of a hundred.

Reviewing a hundred clips without watching a hundred clips

• Scan the grid, not the clips. Wrong exposure, wrong crop and wrong subject are all visible from a poster frame, and that is most of what goes wrong.
• Watch only the ones with hands or faces. Those are the ones where the failure is in the middle of the clip rather than in the first frame.
• Re-run rejects as a smaller batch with one clause changed, not one at a time.
• Keep the constant. If a re-run needs a different constant, it is a different set, and it will look like one.

Done this way a hundred clips is roughly fifteen minutes of attention: a minute writing the constant, three clips of proof, a paste, a scan, and one re-run. The failure mode is never the machinery. It is deciding what varies after the run instead of before it.