Graph Neural Cellular Automata

Every node runs the same learned rule and talks only to its graph neighbors. From a single seed it grows the target pattern — and heals if you drag across it.

View

Single channel

Display

Target ghost Graph edges Glow

Speed 2 steps / frame

Brush radius 0.08

Topology

Surgery on the graph itself — the rule keeps running on whatever remains.

Seed node history

Each row is one of the seed node’s 16 channels, time running left to right. It settles into a stable chemical fingerprint.

The rule’s weights

W₁ — how the 128 hidden units (columns) read the 48 inputs: own state, neighbor mean, difference (row blocks).

W₂ — how those hidden units write the 16 state updates; top block is rgbα.

space pause · r reseed · k clear · v view · e cut edges · E restore
after Distill’s Growing Neural CA — runs entirely in your browser