Adam Hewgill and Brian J Ross.
"Procedural 3D Texture Synthesis Using Genetic Programming".
Computers and Graphics Journal, vol.28, n.4, 2004, pp.569-584.
Preprint (PDF).
Journal web site.
Example A. The following images show procedural textures trained from a model consisting of mountains with white peaks, grey cliffs, and green valley. After extracting sample points from the model, the gradient and blue channels were swapped. The results exhibit a variety of interesting mineral-like effects. The first 2 terrain images in the top row of each solution set are training results. All the solutions use the same experiment parameters, except for different random number seeds.
Example B. These images are trained from a blue cube with coloured 'ribbons' around each orthogonal direction. Vertex normals are shared (interpolated for surrounding faces). This will give a warping effect to results.