Thirty frames of the same jump, and one of them is the one. Your eye knows in a second and can't say why.

Sharpness, it turns out, is a pixel disagreeing with its neighbours. In a sharp picture next-door pixels argue loudly: an edge is light here, dark there, nothing in between. Blur is a truce; everything agrees to be roughly like the pixel beside it. So each frame is shrunk to 512 pixels on its long side (enough to tell blurred from sharp, and a thirty-frame burst reads in well under a second), every pixel is compared with the four around it, and how much arguing there is across the frame is the score.

There's no absolute scale. A cloudless sky has no edges to argue over however good the lens, and a brick wall has thousands. So the score only means anything inside one burst, where every frame looks at the same thing: each is measured against the sharpest of its siblings.

Faces change the rules. If any frame has one, Vision rates how well each face was caught, and that counts for half the score. A frame without a face scores nothing for faces, because a sharp picture of the back of someone's head is a sharp picture of the wrong thing.

The best frame is suggested, with any within two per cent of it, since that close is a coin toss. Nothing is thrown away; the frames are only preselected, and you still choose. It's a jury of pixels, and a jury only gets to recommend.
