Real background plates

Backgrounds were the number one tell in our output, and the reason was structural: the room was being invented twice. Once by the image model, then again by the video model. Two rounds of invention compound, and rooms are where people's eyes go when they're deciding whether something is real.

The fix was to stop inventing them. Composite the person into an actual photograph of an actual room.

Why modified beats faithful

Counterintuitively, you shouldn't keep the photo as-is. Give the model a change list — "make it a different version of this room" — and the result reads as MORE real than the faithful composite, not less.

The reason is that the model keeps the photograph's structure, lighting physics and clutter statistics while executing your changes. You inherit everything that's hard to fake and lose the thing you don't want, which is recognisability.

The one that prompting couldn't fix

Plates whose implied camera position is arm's-length — bedrooms especially — make the model render an extended arm holding the camera. We tried both directions: negative phrasing ("no arm extended") and positive pose language ("hands resting on knees, both hands visible"). Both failed. The plate's camera geometry wins over the prompt.

What worked was a targeted edit pass afterwards: "remove the bare arm reaching into the foreground, keep everything else exactly the same." Surgical, no collateral change. Worth knowing as a general lesson — when geometry is fighting you, fix it downstream instead of arguing with the prompt.