Five natural patterns, each learned by a tiny neural cellular automaton. Choose two, move the slider, and a model trained to hold both renders the blend in real time.
About this room
Choose a pattern on the left and one on the right, then move the slider to blend between them in real time. The shadows and highlights pickers colour the output.
The room shows multi-target training (Method 1 of three in the report). This is the only one of the methods that runs interactively: a single trained model takes the mixing coordinate m as a direct input, so any value of m renders instantly as the network runs. In the other two methods, m selects between combinations of pre-computed weights, which are generated ahead of time rather than at the slider. Their outputs appear in the gallery. The palette here is a simple two-stop gradient; the gallery works use the fuller pipeline described in the report (four colour stops, tonal inversion, gamma, and grain).
Exhibition
The realised works: hybrids that succeeded, and some that failed in interesting ways, each labelled with its parents, method, and mixing value.