As the author says, the script is rather adhoc. The script errors out for me, but from the comments, the basic procedure is to break up a low-resolution versio

As the author says, the script is rather adhoc. The script errors out for me, but from the comments, the basic procedure is to break up a low-resolution version of the image into blocks, and then cluster these blocks on some criterion. Let me see if I can cook up something for you after work :) The simplest speedup I can see is to memoize the distance function calls.