Neural cellular automata grow a full-res image from purely local rules

Cells that only know their neighbors, now at full resolution

Cells to Pixels, a SIGGRAPH 2026 paper from EPFL and Google Research, tackles the constraint that has kept neural cellular automata (NCAs) as a beautiful toy. An NCA is a grid of identical cells, each running the same learned local rule, talking only to its immediate neighbors, until the grid self-organizes into a target pattern. Like living tissue, it can regenerate after damage and resist perturbation. The price of that strict locality was resolution. Memory and compute scale quadratically with grid size, so NCAs stayed low-res.

The paper's fix is a hybrid. The NCA runs on a coarse grid, and a lightweight implicit decoder — a small shared network mapping a cell's local state plus a coordinate to color and surface attributes — renders the fine pixels on top. Both halves stay local and parallel, so the model produces high-resolution output in real time while keeping the self-organizing behavior that makes NCAs interesting. It trains end to end against target images, with separate losses for morphogenesis (growing from a seed) and texture synthesis.

The demos are interactive, and people immediately broke them

The page ships live demos (a PBR texture sandbox, a growing demo, a 3D-texture demo on mesh geometry), and the thread treated them as a stress test. The brush mode turned out to be fragile:

"The automata just completely destroys the image if I draw too much over the stabilized image with the brush. 5 horizontal swipes are enough to destroy the kitty, is that to be expected?" — a developer who filmed the failure

That is arguably the system working as designed — perturb a self-organizing system hard enough and it loses the basin of attraction — but it reads as a bug to a first-time user. Others wanted to push it the other way: stamp a kitten on top of a chameleon instead of resetting the canvas on every target switch, or simply grow a second centipede for extra points.

"It's memorizing enough to do texture sampling with only local information"

The clearest plain-English explanation in the thread came from a reader reasoning out loud about what the model actually does:

"When creating each pixel, there's no coordinate to look up, instead it's using only a set of rules like Conway's game of life. But the rules come from a neural network trained on the image, so… it's kind of memorizing enough information to effectively do the same thing as texture sampling, but using only local information." — a commenter working it out from first principles

For anyone lost by the abstract — and a few said they were — a high-karma reader pointed to the lineage. This group's earlier, gentler write-up is the 2020 Distill article *Growing Neural Cellular Automata*, which a commenter recommended as the better on-ramp. One speculative thread wondered whether the regeneration property could jump domains entirely — a bio-inspired Kubernetes that heals itself when nodes fail.

A self-organizing system finally usable at the resolution people see

The research line has always been about robustness over fidelity: systems that rebuild themselves from local rules, the way an organism does. What kept it academic was that the output looked it. By splitting the work between a coarse self-organizing grid and a cheap decoder for the pixels, the paper keeps the regeneration property while delivering output at the resolution a person actually looks at. The brittle brush in the demo is the honest part: a system that heals from damage is, by the same math, a system you can shove out of its stable state if you push hard enough.