It seems to use characters as pixels. While the effect is interesting, it could be even better if the algorithm was aware of character shape and choose the best character to represent pixels underneath it (for example by using the least squares method with some additional filters).
