Take a movie, capture all faces in all frames. Let a "faceframe" be a frame of the movie scaled and rotated and translated and restricted to a rectangle so that the eyes are in some canonical position and the rectangle is roughly the size of the whole face.
Run a PCA on all faceframes. This gives an axis of maximal variation. Sort all faceframes by their value projected onto this axis. Display that as an animation.
I wonder if you'd actually end up with a very good clustering that tends to put the same person's face (in different orientations) in a contiguous section of the resulting animation. Still would be neat to watch. Maybe you'd want to do a highpass on each frame before the PCA to make sure you don't end up with just a ramp from "darkest ambient lighting" to "lightest ambient lighting".
