All the synthesized textures shown in our paper are relatively small compared to the original samples. We do this in order to show the most amount of details of the textures, and it will be pretty impractical to show extremely large generated textures. Some people are interested to know if there will be any "periodicity" in the output if it is much larger than the input.

Let me use my favorite texture as the test example. It is very small, with size 64x64 only.

And here are the synthesized results with size 512x512. The first one is generated using the "baseline" algorithm using only 1 pass. The second one is generated by the two pass algorithm that we used for constrained synthesis. All textures are generated using 4-level Gaussian pyramids, with neighorhood sizes {9x9,2} at each level of the pyramid. (All the symbols used are the same as those in the paper). You can right click on any of them to view the real sized version.

Single Pass

Two Pass

Tile

There seems to be no periodicity in the generated textures, if you compare them size by size with a verbatim tiling one. Actually this shouldn't be a surprise, since we start the textures as random noises and these random pixels will destroy most (if not all) tendencies for replication. However, you can still see some "pseudo-periodicities" in parts of the textures (similar to the Penrose tiling). This should be the case, since this is caused by the local similarity between input and output textures.

And, if you are not happy enough, here are even larger results, with size 1024x1024.