Start with your face.

What is actually happening

The plate is not an image. It is — individual particles, each carrying one colour, and the picture you see is only where they happen to be standing.

1. The assignment. The target is chopped into the same number of slots. Every particle has to be given exactly one slot — a bijection — and we want the one that minimises total colour error. That is an optimal‑transport problem. It is solved live: a luminance rank‑match warm start, then a few hundred thousand swap proposals per second, each accepted only if it lowers the total cost. Proposals are drawn from a colour‑bucketed index of the target, so the solver spends its time on swaps that could plausibly help.
2. The metric. Colour distance is measured in Oklab, so "closer" means closer to the eye rather than closer in RGB. Coherence adds a squared‑distance penalty, which makes particles prefer a decent nearby slot over a perfect one across the plate.
3. The transport. Nobody teleports. Each particle's desire to reach its slot is splatted into a velocity grid, that field is made divergence‑free with a multigrid pressure solve, and the particles are advected through the result. Incompressibility is why it swirls instead of collapsing — the field physically cannot let everyone take the straight line.
4. Your paint. Painting overwrites particle colours, which changes the optimal assignment, which the solver notices within a frame or two. Fresh strokes hold — they resist the pull, and then surrender at whatever rate you choose. Pigment drift decides whether an arrived particle may take on its slot's colour. At zero, only colours you have actually put on the plate can ever appear.

In the toy both dials follow a very slow build — drift runs 0.03 px/s at one second, 0.06 at six, 1.4 at twelve and 18 at thirty — with a per‑particle lead of up to seven seconds set by a noise‑warped diagonal, so the plate wakes a region at a time rather than all at once. In here there is no schedule and you drive.