Take a look at this quick demonstration of a pretty nice color reduction method I created while experimenting with clustering algorithms in C++. The algorithm clusters a graph of connected points inside of the color space of an image. The points "communicate" along their edges (pink lines) to better locate themselves and their neighbors over time. New nodes are also added one at a time on the edges where there is a large amount of error (which means the nodes on that edge are doing a poor job fitting the color space). It is a computationally expensive algorithm in itself, and I'm working on a faster trimmed down version.