Model renderingInput video rendered along the reference camera path
COMPARED WITH
Held-out recordingWhat the synchronized reference crop actually captured
PermanenceIs the target present?
ContinuityIs it in the recorded place?
AppearanceDoes it still look like itself?