Three test segments before you commit a season
A season is a long commitment to one set of picture behaviour. Three segments taken from the storyboard you have already written, run on every candidate with the same references, settle in an afternoon what a month of production would otherwise reveal. As of 2026-09-12.
| Check | Pass rule |
|---|---|
| Camera | The move requested is the move performed, and the framing holds for the whole shot |
| Cast | Nobody extra in frame, nobody missing |
| Dialogue fit | The line finishes inside the shot, with the mouth closed before the cut |
| Face and costume | Matches the character sheet in close-up and in the wide |
| Hands and contact | Objects are held, passed and put down without changing hands by themselves |
| Speaker | The line comes from the face that was named |
Inclusion rule. Checks that can be judged from a single viewing without measurement. Anything needing frame-by-frame inspection belongs to a later review, not to a candidate test. Order. Alphabetical by check.
1Use your own material, not the demo
A showreel is a selection, and it was selected by someone who knew what the model does well. It is evidence that something is possible once, which is a different claim from the one a season needs.
Three segments from the storyboard already written, with the same references attached, produce evidence about the material that will actually be generated. Keep the three fixed so candidates are comparable.
2Pick the three segments to be awkward
One two-hander with dialogue, because speaker attribution is where most takes fail. One with a physical hand-off or a contact, because that is where invented movement appears. One wide establishing shot with the cast present, because that is where faces drift furthest from the sheets.
An easy three produces a flattering result from every candidate and settles nothing. The point of the exercise is to find the difference, which means testing where the differences live.
3Run each segment more than once
A single take shows what a model can do; three takes show what it does reliably, which is the property a season depends on. A candidate that passes once in three is not cheaper than one that passes twice in three, whatever the rate says.
Recording the pass count per check turns the comparison into a small table instead of an argument, and it is the same tally the take log keeps during production.
4Cost the season before judging the picture
Put the delivered duration of the whole season against each candidate at the rate and resolution you would actually ship at. Some candidates are eliminated by that arithmetic before any picture is looked at, which saves testing them.
Doing it in the other order is how productions fall for a model they cannot afford to finish with, and discover it in episode twelve.
5Write down what the test did not cover
Three segments cannot tell you about queue behaviour under load, about how the channel handles a refusal, or about what happens when a model version changes mid-season. Those are real risks and the test is silent on all three.
Naming them is what keeps a favourable test from being read as a guarantee. The test decides which candidates are plausible; the first episode decides which one is right.
Craft notes, not a register entry. No line here is attributed to a tool, and none should be read as a claim about one. The sourced material is on the register. Related: Mixing models, A blank cell.
6Continuity in practice
Serial work like “Testing a candidate” depends on continuity; SceneMixer's guides describe how it keeps it:
The cast is generated once, at the start, as full-body reference sheets. Every segment that features a character is generated against that character's sheet, with face and body crops attached as references — so the face in episode 1 is the face in episode 40.
SceneMixer, AI Short Drama Generator — “How does it keep the same actor across episodes?” (read 2026-10-13)
The lines are re-heard and written in the language your project is set to, and the finished episode is spoken by the voices attached to your own characters.
SceneMixer, How to Remake a Video with AI — “Can I remake something in a different language?” (read 2026-10-13)