Neural cellular automata: grow, persist, regenerate

Three identical tiny networks (~4.7k parameters) were trained to grow this ladybug from a single seed cell — each under a different training regime. Every cell repeatedly applies the same learned rule using only its 3×3 neighbourhood; everything you see is live inference.

Draw on any panel to erase cells and watch what each model does about it.

target

40×40 px

step 0

speed seed

brush mode damage all panels together

What to look for. All three models grow the ladybug. The differences appear over time and under attack: grow-only was always trained from the seed for ~50 steps, so it has never seen its own grown pattern — left running, it typically drifts or decays, and it has no reason to heal. + persist also trained on its own previously-grown states (a sample pool), so it learns to hold the pattern steady — but damage is still foreign to it. + regenerate additionally had random holes punched into training states, so erased regions regrow. The half-cut and your free-hand scribbles are damage no model ever saw in training. Each panel shows the pixel error (MSE) against the target — lower is better, and it is recomputed live.

A riff on “Growing Neural Cellular Automata” (Mordvintsev et al., distill.pub 2020) — independent re-implementation, run as a controlled training-regime comparison; the steerable panel is our addition. Details & code: REPORT.md in the accompanying repository.