Water can't be squashed, so wherever you push it, it has to make room by flowing somewhere else. The tank is divided into a grid of 150 by 120 cells, each with a velocity. Every step, the water carries itself along (each cell takes the velocity of whatever was upstream of it), then the model works out the pressure needed to push the water back to the same density everywhere, and takes it away. What's left is a flow that never piles up or thins out anywhere, just like a real liquid.

The clip shows the Cylinder scene: a current flows past a round post, and swirls peel off each side in turn, a pattern called a von Kármán vortex street. The same thing happens to wind blowing past chimneys, and to clouds downwind of islands.

Turn on a Current, and the left edge becomes an inlet where water flows in at a steady speed, releasing streaks of dye, like smoke in a wind tunnel. The right edge is an outlet, which calms the water down before it wraps round to the inlet again. In Two slits, the current squeezes through two gaps as a pair of jets, which wobble, break up and mix.

Moving dye around a grid tends to blur it, a little more every step. To keep the streaks sharp, the model uses MacCormack's method: it moves the dye forward a step, then back again, and whatever didn't survive the round trip shows how much blurring it made, so it can take that back off.

Working out the pressure exactly would take far too long for every step, so the model improves its guess a set number of times instead: 16, which is plenty to look right. With only a couple of tries the water would be noticeably squashy, and swirls would fade quickly.

The original Java version stored each cell's velocity at its centre, as Stam's paper does. That's simple, but it leaves a checkerboard pattern that the pressure can't see, and so can't remove. This version stores the flow through each cell's edges instead (a staggered grid), which the pressure can balance exactly.