He supports different probabilities for the tiles. Though he doesn't, you could vary those probabilities spatially. You could even apply the algorithm with one tileset to get Image #1, then scale up Image #1 and use its colors to determine the probabilities for a second application of the algorithm for Image #2, possibly with a different tileset. Recurse as desired.
On the other hand, it might turn out that when you do things that way, you don't need the edge-matching rule at all; you can just say which colors can correspond to which tiles at the next level of resolution down, maybe with probability, and rely on the high-level tiles to guarantee continuity.
On the other hand, it might turn out that when you do things that way, you don't need the edge-matching rule at all; you can just say which colors can correspond to which tiles at the next level of resolution down, maybe with probability, and rely on the high-level tiles to guarantee continuity.