Motion vs averaging

How drift changes what a frame average recovers. Averaging a run of frames cancels the noise and resolves the mean; drift moves the value between frames, so the average smears instead. This page measures how much: what a naive frame average gets back, and what one that undoes the motion first gets back.

Motion Amplitude 12px Independent regions 1 Frames captured 64

What you see. The block is moving; the grid behind it is not.

One screenshot

Naive average

Motion recovered

Pick a motion and press the button.

readability is Pearson r against the true value 1.0 = fully readable ~0.1 = noise

What the three panels are

One screenshot is a single captured frame -- what a still capture or a phone photo gets. It should always be noise; that is the base defence and motion has nothing to do with it.
Naive average is the lazy recording attack: average every frame as recorded. With no motion it reads the value straight back. With motion it smears, because the value is no longer on the same pixels.
Motion recovered is the honest attacker: this page knows the exact motion it applied, so it shifts every frame back before averaging -- which is what a competent attacker's motion estimation aims to do. This is the ceiling of what motion can be made to give up.

What to try

static -- the baseline. With nothing moving, the naive and motion-aware averages agree, and every motion setting below is read against this.
sine, amplitude 12 -- naive smears a little, motion-recovered snaps back to fully readable. A smooth drift is two numbers to guess.
random, amplitude 12+ -- naive smears hard. Motion-recovered still wins, because the value is fully present once each frame is placed. Random only removes the shortcut of a formula.
regions 6, random -- the block breaks into independently moving pieces, closest to per-word motion. Hardest naive result, same recovered result.

The recovered panel is the point. However hard the naive average smears, the value is still fully there -- motion hides where it is, not whether it is, and the value is the high-contrast pattern an attacker locks onto to undo the motion. This is why nocap ships drift as a visual on the promo and makes no protection claim for it. The algorithms page has the measured attack table.

The honest gap. The middle panel is a naive attacker and the right is one who recovers the motion perfectly. A real attacker lands between them, depending on how good their motion estimation is -- this page does not ship that estimator, so it shows the two bounds rather than pretending to know where a given attacker falls. And the motion smooth enough to read comfortably is the motion easiest to recover: the same single budget as everything else.