Cells2Pixels gives neural cellular automata an 8x resolution path

Why it matters

Cells2Pixels attacks the main practical limit on neural cellular automata: high-resolution output without blowing up the cellular grid, memory use or step count.

Ehsan Pajouheshgar (@Esychology) used a July 9th thread on X to put a practical claim around Cells2Pixels, his SIGGRAPH 2026 work on neural cellular automata: the model can deliver what he described as an approximately 8x resolution boost with minimal extra parameters by treating cells as local neural fields instead of pixels.

https://x.com/Esychology/status/2075259224255349234

The project, formally titled Neural Cellular Automata: From Cells to Pixels, comes from Pajouheshgar, Yitao Xu (@xuyitao5), Ali Abbasi, Alexander Mordvintsev (@zzznah), Wenzel Jakob (@wenzeljakob), and Sabine Susstrunk (@ssusstrunk). The project page lists EPFL as the affiliation for Pajouheshgar, Xu, Abbasi, Jakob and Susstrunk, and Google Research as Mordvintsev's affiliation. Abbasi's contribution is marked as work done during an EPFL internship.

Pajouheshgar is a postdoctoral researcher at EPFL whose published work has stayed close to the same question: how simple local update rules can produce coordinated visual behavior. His personal site says he did his PhD at EPFL under Susstrunk, studied computer engineering with a math minor at Sharif University of Technology, and works across self-organizing systems, computer vision, computer graphics and artificial life. Cells2Pixels follows a run of related papers from the same circle, including Mesh Neural Cellular Automata at SIGGRAPH 2024 and DyNCA at CVPR 2023.

What changed

Neural cellular automata, or NCAs, are learned systems in which each cell repeatedly updates itself from nearby cell states. That locality is the point: the models can grow shapes from a seed, synthesize textures, repair damaged patterns and generate dynamics without a centralized scene representation. The scaling problem is also local. Pajouheshgar wrote that when cells only communicate with neighbors, larger grids require more steps to form and maintain structure, training becomes unstable, and memory and compute rise quadratically with grid size.

Cells2Pixels keeps the cellular automata grid coarse. Instead of reading each cell as a pixel, the system attaches a Local Pattern Producing Network, or LPPN, to each cell neighborhood. The LPPN takes the cell state and a local coordinate as input, then outputs the color or appearance property at that point. In the authors' phrasing, that decouples render resolution from grid size.

The arXiv paper was first submitted on June 28th, 2025, and the current version was posted on May 1st, 2026. The paper states that the NCA evolves on a coarse grid while a lightweight implicit decoder maps cell states and local coordinates to appearance attributes, allowing outputs to render at arbitrary resolution. The authors also say the model remains parallelizable because both the NCA updates and decoder computations stay local.

That design matters because Cells2Pixels is not trying to scale NCAs by making the cellular grid bigger. It moves the detail into a tiny local neural field while preserving the automata's self-organization dynamics. The practical bet is that high-resolution textures, morphologies and mesh appearances can be generated from a smaller simulation state, rather than paying the full memory and step-count cost of a dense cellular grid.

The demo and code

The team published online demos and says the models run fully on-device. The demo page includes interactive examples for PBR texture synthesis, growth, 3D texture and a bonus dynamic texture mode, and invites users to interact directly with the canvas. The project page describes evaluation across 2D and 3D grids as well as mesh domains.

The GitHub repository is the official implementation of the SIGGRAPH 2026 paper. It includes configurations for 2D growth, PBR texture synthesis, mesh texture synthesis and 3D texture generation, plus exploratory modes for dynamic textures, growing radiance fields and voxel growth. As of the retrieved GitHub page, the repo showed 58 stars, 5 forks and 18 commits.

The repo also makes the compute dependencies concrete. The README pins PyTorch to 2.8.0 and Kaolin to 0.18.0 for mesh-related experiments, with Kaolin needed for mesh rendering and rasterization. The same training entry point is used across experiments, with separate YAML configs for the different modes.

Why the result matters outside graphics research

Cells2Pixels lands in a period when generative systems are dominated by large diffusion and transformer architectures, yet many interactive applications still need small, local, responsive models. NCAs sit on the other side of that trade: they are compact, spatial and inherently iterative. Their weakness has been scale. A model that can keep the automata coarse while rendering detail at higher resolution gives the NCA line of work a cleaner path into real-time textures, editable simulations and browser-native visual systems.

The open question is how far the local-field trick travels beyond the paper's chosen visual domains. The team reports experiments on growing morphologies and texture synthesis in 2D, 3D and meshes. That is meaningful coverage for graphics, but it does not by itself prove that NCAs can become a general substitute for higher-capacity generative models in production art tools or simulation engines.

For Pajouheshgar and his coauthors, the sharper claim is architectural. Cells2Pixels argues that the bottleneck was the pixel-level readout, rather than the cellular automata premise itself. If that holds across more demanding scenes, the old limitation of neural CAs as low-resolution curiosities starts to look like an implementation constraint that can be engineered around.