Reference remix take kinds: which model renders each shot

A remake is a list of takes

When Sume's agent remakes a reference video shot for shot with your product, it does not render one long clip. It writes a plan of takes, and every take has a kind. The kind decides which model makes it. The list of kinds lives in the repository's contract package: talking, performed, voice, insert, still, recast, genjutsu and product swap.

The map matters because each model has a narrow job, and the wrong pairing is the usual way a remake drifts into a different video.

The map

Why the take is measured after it renders

The instructions say the remake's length and every cut come from the new voice, not the reference's seconds. After a take exists, a reference_take call transcribes that take's own audio and aligns it with its script, so each word has the frame it is actually spoken on. A reference_timeline call then lays the takes out so a cut sits on a word or a take edge.

That is why a recast take is special. A Recast clip keeps the source's own camera, so the instructions say that when one of these source-footage models remakes the whole clip, the clip it returns is the program, and only the music bed and any overlay go over it.

What this means for you

You do not pick take kinds by hand. You point at a reference and say what changes. If the change is the people and the footage is one take, expect a recast take. If the reference is a talking head and you have a new script, expect a talking take. Each model's limits still apply: Recast needs 5 to 30 seconds with no shot over 15, and the Video Router docs list the other rows' windows.