Critically Thomas: Growing Neural Cellular Automata

This repository adapts the neural cellular automata (NCA) training method from Growing Neural Cellular Automata by Mordvintsev, Randazzo, Niklasson, and Levin (Distill, 2020). Instead of the original emoji target, the checked-in experiment trains a 16-channel cellular automaton to grow the Critically Thomas wordmark from a single live cell.

The model sees each cell and its Sobel-filtered neighbourhood, applies a small learned update network, and randomly updates roughly half of the cells at each step. A pool of intermediate states exposes the model to different stages of growth during training. result.gif is a generated output from this project; it is included as a visual record, not as a benchmark.

Run the experiment

The project uses Python 3.13 and uv:

uv sync --locked
uv run critically-thomas-nca

The entry point runs title-cellular-automata/nca.py. On its first run it trains from scratch, writes a model checkpoint to title-cellular-automata/nca.pt, and renders result.gif. If that checkpoint already exists, it is loaded and training continues before the GIF is regenerated.

Training is compute-intensive. The script selects CUDA first, then Apple MPS, and falls back to CPU. CUDA requires a working NVIDIA driver compatible with the PyTorch build installed by uv. MPS requires a supported Apple-silicon Mac and macOS/PyTorch combination. CPU execution needs no accelerator, but the 10,000-step default can take considerably longer. The repository does not include a trained checkpoint.

Repository contents

• title-cellular-automata/nca.py contains the checked-in 1× target training and rendering experiment.
• title-cellular-automata/target.npy is the RGBA training target; target.png and target@11x.png are viewable versions.
• result.gif is a checked-in generated result.
• main.py provides the installed command-line entry point.

Attribution

The architecture, perception filters, stochastic cell updates, alive-mask logic, and training-pool approach are adapted from the Distill article and its accompanying implementation:

Mordvintsev, A., Randazzo, E., Niklasson, E., & Levin, M. (2020). Growing Neural Cellular Automata. Distill. https://doi.org/10.23915/distill.00023

The target artwork and experiment-specific changes are part of this project. No license is asserted here; consult the upstream article and implementation for the terms that apply to their material.