Researchers push neural cellular automata beyond low-res grids

Ehsan Pajouheshgar, Yitao Xu, Ali Abbasi, Alexander Mordvintsev, Wenzel Jakob, and Sabine Süsstrunk describe a way to scale neural cellular automata in Neural Cellular Automata: From Cells to Pixels, a June 2025 arXiv paper.

The authors target a core limit in neural cellular automata, or NCAs. Researchers use NCAs to train grids of identical cells, each cell updates itself from nearby state, and the full system grows patterns from local rules. That setup gives NCAs their appeal: a model can grow an image from a seed, repair damage, and produce texture dynamics without a global controller.

High resolution exposes the cost. A bigger grid gives the model more cells to update, more memory to track, and more training steps to move information across space. The authors name three bottlenecks: grid cost grows with image area, local updates slow long-range coordination, and real-time inference strains hardware at dense resolution.

Their answer separates growth from rendering. The NCA runs on a coarse grid, where cells can coordinate at lower cost. A small implicit decoder then maps each cell state plus local coordinates into visual attributes. In practice, the NCA learns the structure of the object or texture, and the decoder draws the pixels.

That split matters because the authors keep both pieces local. Each NCA update reads nearby cells. Each decoder query reads the state around a point. GPUs can run those operations in parallel, so the model can render more pixels without asking the NCA to simulate a full high-res grid.

The paper covers morphogenesis and texture synthesis. In morphogenesis tasks, the model grows an output from a seed. In texture tasks, the model produces repeatable visual detail. The authors add task-specific losses so the model can train against high-res targets without carrying the full high-res simulation through each update.

The authors also test the idea beyond flat images. They report experiments on 2D grids, 3D grids, and mesh domains. That range gives the method a stronger claim than a texture trick. The same coarse-state, local-decoder design can work across domains where researchers want compact self-organizing behavior and high-res output.

The paper does not announce a company, funding round, or investors. Its market relevance sits in rendering and generative systems, where developers want real-time high-res outputs without large inference costs. Game engines, simulation tools, procedural asset systems, and design software could use this kind of model if future work turns the research code into production tools.

The trade-off comes from the same design choice that gives the method speed. A coarse NCA may miss fine causal structure that a dense NCA could model. The decoder can add high-res appearance, but the coarse grid controls the underlying state. Developers would need to test whether the method preserves behavior under edits, damage, animation, and domain changes.

The paper advances NCAs by treating pixels as a rendering target rather than the simulation substrate. That choice lets researchers keep the self-organizing character of NCAs while moving the output resolution closer to production needs.