Keeping the room, swapping the faces

"Keep the room and swap the faces" sounds like one request. It is really two, and they are in tension. Half the inputs are there to be copied exactly — the room, the camera, the timing, the movement. The other half are there to be substituted completely — the two people.

When the instruction blurs the two, the model picks the easier reading. Ask it to "make a video of two people singing in an orange room" and you have asked it to invent everything, including the people. It will happily do that, and the result will look nothing like the people you uploaded.

The fix is to state the roles explicitly and early: this image is the person on the left, that image is the person on the right, these two are the cast, and nobody else appears. It sounds redundant. It is the single highest-value sentence in the instruction.

On runs where that sentence was missing, the model reproduced the performers from the reference footage, down to their outfits, and ignored the uploaded photos entirely. Same inputs, same model, same everything else — the only difference was whether the instruction said whose faces belonged in the result.

The same logic applies to the room, and it bites in a more embarrassing way. An element the instruction does not mention does not stay as it was — it gets invented. Our own first attempt at the scene never mentioned the microphone, and produced a perfectly plausible microphone on a floor stand: exactly the wrong object for a format whose whole recognisability rests on a mic hanging from the ceiling.

So the set list is written out: the orange walls and ceiling, the lamp, the counter, the single microphone hanging on its cable, the lighting. Boring to write, and it is the difference between a clip that reads as the trend and a clip that reads as vaguely similar.

The rendered clip arrives with an audio track of its own. That track is not what the finished video should carry, so it is replaced — the video is kept exactly as rendered and the intended soundtrack is attached in its place.

That separation is also what keeps the reference footage usable. Feeding a reference clip that already carries a copyrighted recording is what gets jobs rejected by some providers; strip the audio first, attach the intended track at delivery, and the render never depends on a recording you do not control.

This class of model follows a performance rather than cloning it. It will match the gestures, the camera and the timing closely enough that the clip is instantly recognisable. It will not give you a frame-accurate replica, and no amount of prompt wording will change that.

What you can control is whether it is the right room, the right two faces, and the right moment. That is what the instruction is built around.