Show HN: High-Resolution Neural Cellular Automata

Existing Neural Cellular Automata (NCA) excel at generating complex patterns, but they have a limitation in that as the grid size increases, training time and memory requirements grow quadratically, making real-time inference difficult.

To address this, this research proposes a hybrid model that combines an NCA evolving on a Coarse Grid with a lightweight Implicit Decoder (LPPN) that maps cell states and local coordinates to visual attributes.

The key features are as follows:

• Arbitrary Resolution Support: Since the decoder uses local coordinates, high-resolution rendering is possible regardless of output resolution.
• Efficient Parallel Processing: Both NCA updates and decoder computations are performed locally, maintaining high parallelism and enabling real-time inference.
• Application Across Various Domains: Experiments on 2D/3D grid and mesh domains demonstrated morphogenesis and texture synthesis capabilities.
• Efficient Training: By introducing task-specific losses for morphogenesis and texture synthesis, high-resolution outputs can be learned without additional memory overhead.