NeuralCA.js

/** @license 
  * Copyright 2023 Google LLC.
  * SPDX-License-Identifier: Apache-2.0 
  */

// Based on "μNCA: Texture Generation with Ultra-Compact Neural Cellular Automata"
// https://arxiv.org/abs/2111.13545
class NeuralCA {
    static Tags = ['2d', 'ca'];
    constructor() {
      this.W = new Float32Array([
        -67,1,2,44,-13,-59,4,30,-1,16,-57,9,-10,-4,-2,-41,    19,-18,-1,8,-4,35,8,0,-4,-4,-1,0,34,31,21,-25,
        4,13,18,-57,-79,-22,-25,71,-12,-11,24,27,-17,-8,-7,6, 11,10,4,0,4,1,2,7,-26,-33,-15,-3,22,27,20,-34]);
      this.b = new Float32Array([2,-5,-14,9]);
    }
    frame(glsl, {DPR}) {
        const state = glsl({W:this.W, b:this.b, FP:`
        uniform mat4 W[4];
        uniform vec4 b;
        vec4 rule(vec4 s, vec4 p) {
          return 1e-3*(b + W[0]*s + W[1]*p + W[2]*abs(s) + W[3]*abs(p));     
        }
        void fragment() {
          vec4 s = Src(UV);
          if (s == vec4(0)) {
            ivec2 I = ivec2(gl_FragCoord.xy);
            FOut = 0.1+vec4(hash(I.xyy).x)*0.4;
            return;
          }
          vec2 dp = Src_step();
          float x=UV.x, y=UV.y;
          float l=x-dp.x, r=x+dp.x, u=y-dp.y, d=y+dp.y;
          #define R(x, y) Src(vec2(x, y))
          // perception
          vec4 p = R(l,u)*vec4(1,1,-1, 1) + R(x,u)*vec4(2,2,0, 2) + R(r,u)*vec4(1,1,1, 1)
                 + R(l,y)*vec4(2,2,-2, 0) +  s*vec4(-12,-12,0, 0) + R(r,y)*vec4(2,2,2, 0)
                 + R(l,d)*vec4(1,1,-1,-1) + R(x,d)*vec4(2,2,0,-2) + R(r,d)*vec4(1,1,1,-1);
          vec4 ds = rule(s-0.5, p);  // NCA rule application
          FOut = s+ds;
        }`}, {story:2, scale:1/4/DPR, tag:'state'});
        glsl({tex:state[0], FP:`tex(UV)*2.-.5`});
    }
}