If you put the UVs in world space (and adjust UV scales accordingly) then it should tile across multiple terrains correctly.

Actually my version is not the same as the one in the paper and samples unity released. They do their blending by trying to preserve a histogram of the original image. The advantage of this is that it will work for any texture, however, there are many disadvantages as well: - Several minutes of texture preprocessing time - You have to store extra data in lookup textures - Dependent texture reads which slow down the shader - A lot more code and math - The blending of the texture is not correlated across textures; ie: The albedo may blend differently than the normal - From my experiments, the blending tends to wash out the textures much more than my technique. My modification to this technique was to keep the basic algorithm for sampling the textures, but remove the histogram preserving blend and replace it with a height map based blend. This removes all the downsides listed above, and in my opinion looks better as well, as the blend are not only correlated between the textures, but also tighter, giving a crisper overall image. Overall, I'm very happy with the result and think it's a nice improvement on the original paper (and have to thank unity for sparking the whole thing in the first place)

Yes, noise size is key with the multi-texture clustering modes! I've actually found that using this technique several smaller textures can look way better than larger ones, because of the rich variance you can get from mixing several texture variations.

Thanks for the review! After this was posted, the MicroSplat texture count limit was raised to 32, making the maximum texture count 96 texture with clustering. (the Advanced details module, which was not created by me, adds another 16).