June 17, 2026
Cells gone wild
Show HN: High-Res Neural Cellular Automata
AI pixel creatures wow the crowd — but commenters are already breaking, questioning, and remixing them
TLDR: This project shows a new way for simple AI-driven “cells” to grow detailed images fast, even at very large sizes. Commenters loved the weirdness but immediately turned it into chaos, asking if the cells know what “up” is, demanding centipede upgrades, and accidentally obliterating digital kittens.
A flashy new Show HN demo is making tiny digital “cells” grow into detailed images in real time, and yes, the science is impressive: instead of trying to build every pixel the hard way, the system lets a simpler grid of cells organize itself, then uses a lightweight helper network to fill in the fine detail at any size. In plain English, it’s a smarter way to make self-growing visuals look sharp without melting your computer. But on Hacker News, the real action wasn’t the paper — it was the peanut gallery immediately poking, prodding, and trying to mutate the thing into chaos.
The strongest reactions split into two camps: awed curiosity and playground gremlin energy. One commenter zeroed in on a surprisingly philosophical question: why are the images always upright, and do these little cells somehow know what “up” is? Another instantly treated the demo like a game, suggesting users should grow a longer centipede — or better yet, spawn a second centipede for extra points. That set the tone fast: less “serious lab breakthrough,” more “what happens if I make the bug army bigger?”
Then came the tiny heartbreaks and mini-drama. One user wanted to mash a kitten onto a chameleon and got genuinely sad when the canvas reset, while another reported that just five brush swipes were enough to completely destroy the kitty, complete with an Imgur video. Even the technical skeptics showed up, wondering if the whole “local updates are limiting” claim clashes with the fact that graphics chips are supposedly great at neighbor-only work. In other words: the demo impressed people, but the comments quickly turned into a mix of bug reports, feature requests, existential questions, and chaotic pet-based experimentation — which is exactly the internet at its best.
Key Points
- •The article presents a hybrid approach that combines a coarse-grid Neural Cellular Automaton with a lightweight implicit decoder to produce outputs at arbitrary resolution.
- •It identifies three main barriers to high-resolution NCAs: quadratic growth in training time and memory with grid size, limited long-range communication from strictly local updates, and high real-time inference cost.
- •The decoder maps interpolated local cell states and local coordinates to appearance attributes such as color and surface normal.
- •The method uses task-specific losses for morphogenesis and texture synthesis to supervise high-resolution outputs with minimal extra memory and compute overhead.
- •Experiments across 2D, 3D, and mesh domains are reported to generate high-resolution outputs in real time while preserving NCA self-organizing behavior.