Show HN: High-Res Neural Cellular Automata
3 hours ago
- #neural cellular automata
- #implicit decoder
- #high-resolution rendering
- Neural Cellular Automata (NCAs) are bio-inspired systems where identical cells apply a learned local rule to self-organize into complex patterns.
- Traditional NCAs face limitations in high-resolution outputs due to quadratic scaling of training time/memory with grid size, strictly local information propagation, and heavy compute demands.
- This work introduces a hybrid model pairing an NCA on a coarse grid with a lightweight implicit decoder (Local Pattern Producing Network) to map cell states and local coordinates to appearance attributes.
- The hybrid model enables rendering outputs at arbitrary resolution while maintaining highly parallelizable inference.
- Task-specific losses for morphogenesis and texture synthesis are introduced to supervise high-resolution outputs efficiently with minimal overhead.
- Experiments across 2D/3D grids and mesh domains show the hybrid models produce high-resolution outputs in real-time and preserve NCA self-organizing behavior.