$μ$NCA: Texture Generation with Ultra-Compact Neural Cellular Automata

  • 2021-11-26 15:26:37
  • Alexander Mordvintsev, Eyvind Niklasson
  • 47

Abstract

We study the problem of example-based procedural texture synthesis usinghighly compact models. Given a sample image, we use differentiable programmingto train a generative process, parameterised by a recurrent Neural CellularAutomata (NCA) rule. Contrary to the common belief that neural networks shouldbe significantly over-parameterised, we demonstrate that our model architectureand training procedure allows for representing complex texture patterns usingjust a few hundred learned parameters, making their expressivity comparable tohand-engineered procedural texture generating programs. The smallest modelsfrom the proposed $\mu$NCA family scale down to 68 parameters. When usingquantisation to one byte per parameter, proposed models can be shrunk to a sizerange between 588 and 68 bytes. Implementation of a texture generator that usesthese parameters to produce images is possible with just a few lines of GLSL orC code.

 

Quick Read (beta)

loading the full paper ...