I've made generative art and it involves studying the algorithms, guessing what inputs might produce good output, and generating hundreds of tests and samples, almost all of which are discarded without value. It takes time and skill to predict what'll work, and touch up the result to something presentable.
It's funny the linked article mentions "planet ruled by little castles" specifically because I've also made those [1] using generative methods, though in this case style transfer rather than language visualization.