Neural Cellular Automata
Show HN: High-Res Neural Cellular Automata
Last verified:
Visit Neural Cellular Automata
What is Neural Cellular Automata?
Neural Cellular Automata (NCAs) are bio-inspired dynamical systems where identical cells iteratively apply a learned local update rule to self-organize into complex patterns, exhibiting regeneration, robustness, and spontaneous dynamics. This work overcomes the traditional limitation of NCAs being confined to low-resolution outputs by pairing an NCA that evolves on a coarse grid with a lightweight implicit decoder called LPPN (Local Pattern Producing Network). The LPPN maps cell states and local coordinates to appearance attributes, enabling the same model to render outputs at arbitrary resolution while preserving the characteristic self-organizing behavior of NCAs.
Key features include real-time high-resolution output generation across 2D/3D grids and mesh domains, fully parallelizable inference since both the decoder and NCA updates are local, and task-specific losses for morphogenesis (growth from a seed) and texture synthesis with minimal additional memory overhead. The interactive WebGL demo runs trained models fully on the GPU inside the browser using SwissGL, offering three modes: morphology growth with direct RGBA output, 2D PBR textures with 9-channel maps, and 3D volumetric textures with NeRF-style ray-marching.
This tool is designed for researchers in computer graphics and computational morphology, graphics artists needing procedural texture generation, game developers requiring PBR textures and mesh shading, and anyone interested in self-organizing systems and emergent behavior. The framework supports texture synthesis on unseen meshes at test time without retraining, volumetric 3D texture synthesis, and offers interactive controllability with features like erasing regions or inserting seeds during simulation.
Neural Cellular Automata pricing
Pricing model: Freemium
Free - The interactive WebGL demo at cells2pixels.github.io is completely free to use. The code is available on GitHub and the paper is on arXiv. No paid plans or subscription tiers are mentioned. The tool runs fully on-device in the browser with no server costs or usage limits.
Neural Cellular Automata pros
- Generates high-resolution outputs in real-time despite NCA typically being limited to low resolution
- Decouples dynamics from appearance, allowing arbitrary render resolution independent of NCA lattice size
- Highly parallelizable inference since both decoder and NCA updates are local operations
- Preserves NCA characteristic properties: robustness, regeneration, and self-organizing behavior
- Interactive WebGL demo runs fully on-device in browser with GPU acceleration via SwissGL
- LPPN adds only 20-30% extra parameters to the model, keeping it lightweight
- Training remains memory-efficient and fast because recurrent updates happen at low resolution
- Supports multiple domains: 2D grids, 3D volumetric grids, and mesh surfaces
- Texture synthesis on meshes works on unseen meshes at test time without retraining
- Outputs 9-channel PBR textures including albedo, normal, and HRA (height, roughness, ambient occlusion)
- Interactive controls allow users to erase regions or insert seeds during simulation
- Stable rendering at resolutions different from training, with quality improving at higher resolutions
- Morphing behavior emerges naturally as an emergent property when swapping model parameters at test time
- After severe damage to morphology, system re-grows missing regions and returns to same attractor
- Task-specific loss functions designed for high-resolution outputs improve quality without exhausting memory
Neural Cellular Automata cons
- LPPN conditions only on intra-primitive coordinates and locally interpolated state without access to cells outside the enclosing primitive
- Can produce faint primitive-aligned patch artifacts due to locality limitation of LPPN
- Volumetric rendering at 8x LPPN scale may not run in real-time without a strong GPU
- Training memory and compute requirements still grow quadratically with grid size for the NCA component
- Strictly local propagation of information impedes long-range cell communication
- Sorting barycentric coordinates makes them non-injective, though LPPN compensates with cell state conditioning
- Auto-correlation regularizer for PBR textures is more sensitive to hyperparameters than base multi-scale OT loss
- Morphing behavior is emergent and not explicitly optimized, making transitions less controllable and predictable
Frequently asked questions about Neural Cellular Automata
What is Neural Cellular Automata and how does it work?
Neural Cellular Automata (NCAs) are bio-inspired dynamical systems where identical cells iteratively apply a learned local update rule to self-organize into complex patterns. The NCA operates on a coarse lattice of cells, and a lightweight implicit decoder called LPPN (Local Pattern Producing Network) maps cell states and local coordinates to appearance attributes. The LPPN receives the locally averaged cell state and local coordinate vector as input, then outputs target properties like color and surface normal at each sampling point. The NCA and LPPN are trained jointly and end-to-end.
What resolution can the tool generate?
The tool enables rendering at arbitrary resolution. Experiments demonstrate outputs at 768² for morphology growth, 1024² for PBR-2D textures and mesh textures, and 512² for volumetric 3D textures. The render resolution is decoupled from the NCA lattice size, so you can render at resolutions different from training, and quality generally improves as output resolution increases.
Is the demo free to use?
Yes, the interactive WebGL demo at cells2pixels.github.io is completely free. It runs trained models fully on the GPU inside the browser using SwissGL, with no server costs, subscription fees, or usage limits mentioned.
What are the three demo modes available?
The demo includes three modes: (1) Morphology growth with direct RGBA output showing growth from a seed, (2) 2D PBR textures with 9-channel maps (albedo, normal, HRA) visualized with a simple PBR shader, and (3) 3D volumetric textures using NeRF-style ray-marching through the learned texture field.
Can I use the model on meshes I haven't trained on?
Yes. The tool adopts MeshNCA where cells live on mesh vertices, and the trained model can be applied to unseen meshes at test time without retraining. During training, the model is trained on an icosphere with approximately 40k vertices, but it generalizes to arbitrary meshes afterward.
How real-time is the inference performance?
Inference is highly real-time. On an NVIDIA A100, steps per second for NCA alone range from 550 (vol-3D) to 3700 (PBR-2D). For render plus LPPN, performance ranges from 13 steps/sec (vol-3D) to 425 steps/sec (growing). For mesh rendering, it's 6 steps/sec with end-to-end rasterization or 110 steps/sec when rasterization is cached. The browser demo defaults to smaller LPPN scales (4x for morphology/PBR, 2x for volumetric) for real-time performance on edge devices.
What interactive features are available in the demo?
The demo offers interactive edits including erasing regions or inserting seeds. Additional controls let users adjust simulation speed and the LPPN sampling scale. Users can click or tap the canvas to interact with the NCA directly.
What PBR texture maps does the tool output?
For 2D PBR texture synthesis, the LPPN outputs a 9-channel field corresponding to three maps: albedo (base color), normal (surface orientation), and HRA which includes height, roughness, and ambient occlusion. These maps can be visualized with a physics-based renderer.
Does the model exhibit regeneration after damage?
Yes, the learned rule defines a stable self-organizing process where after severe damage the system re-grows missing regions and returns to the same attractor. This regeneration capability is a characteristic property of NCAs that is preserved in this hybrid model.
Where can I get the code and paper?
The paper is available on arXiv (arXiv:2506.22899v3 [cs.CV]), and the code is available on GitHub. The interactive demo runs at cells2pixels.github.io. The work was published at SIGGRAPH 2026 and authored by researchers from EPFL and Google Research.