Multi-texture synthesis through signal responsive neural cellular automata.

Area of Science:

• Artificial Intelligence
• Computer Science
• Computational Biology

Background:

• Neural cellular automata (NCAs) are effective for procedural texture generation, modeling global patterns from local interactions.
• Current NCA architectures require separate training for each texture, limiting usability in texture synthesis.

Purpose of the Study:

• To enhance the usability of neural cellular automata for texture synthesis.
• To develop a single NCA model capable of generating multiple textures from individual examples.

Main Methods:

• Training a single, compact NCA for the evolution of multiple textures.
• Encoding texture information within each cell's state using a genomic signal.
• Investigating the influence of genomic information and cost functions on NCA evolution.

Main Results:

• The proposed NCA successfully generates multiple textures using a unified model.
• The NCA supports interpolation between learned textures, enabling texture editing.
• Generated textures can merge and coexist within the same automaton.

Conclusions:

• A single neural cellular automaton can be trained to generate and manipulate multiple textures.
• The genomic signal within cell states is key to NCA's texture generation capabilities.
• This work advances NCA applications in procedural content generation and texture synthesis.