The more scientific question behind all this is fractal classification of real landscapes. By computing the fractal exponential factor from topographical data it is possible to classify the world around us depending on scale and location.
Each of these worlds is uniquely defined by one single number: the initial seed of the (here linear congruential) random number generator. By simply using different seeds, a great variety of worlds can be generated. By right clicking on the worlds you can view the images at full resolution. You can theoretically zoom into the terrain infinitely, although you need to create and store a lot of random numbers for this.
The images have been computed and rendered (using OpenGL) on an SGI O2 in a few seconds. Color just depends on the height value, while oceans are simply constructed by thresholding. On some worlds you can see some artifacts from the initial (octahedral) mesh. These could be circumvented by choosing a finer initial mesh, although in the long run they cannot be avoided once you zoom in. Dominant structures will always generally align with a coarser triangulation.