Show HN: High-Res Neural Cellular Automata (cells2pixels.github.io)

A new approach to Neural Cellular Automata (NCAs) has been introduced, enabling high-resolution outputs while maintaining the self-organizing capabilities characteristic of these systems. Traditionally, NCAs have struggled with low-resolution generation due to challenges like increased training costs and limited information propagation across the grid. The solution lies in a hybrid model that simultaneously employs a coarse grid NCA alongside a lightweight implicit decoder that translates cell states into appearance attributes, allowing for arbitrary resolution rendering. This innovative framework remains efficient due to its local processing nature, which enhances parallelization during inference. The significance of this advancement for the AI/ML community is profound, particularly in applications like texture synthesis and morphogenesis, where high-resolution results are crucial. By integrating task-specific losses for both growth from a seed and texture generation, the model efficiently supervises outputs while minimizing memory and computational demands. Experiments demonstrate successful application across 2D/3D grids and mesh domains, revealing the potential of this method to revolutionize how NCAs can be used in real-time applications without sacrificing quality or performance.