I recently decided to try a program called PixPlant which claimed it could generate seamless textures from photos, so I tried it on some somewhat and very complex images of mine and the algorithm completely failed to make a nice enough texture.
It uses an imagesynth algorithm which, which did poorly for all of the images I tried except one or two. The image elements repeated too often and I could tell it didn’t extract irregular shaped elements which would hold such an algorithm back. Even with my simple image of hundreds of circles, I noticed some circles were cut off.
So begs the question, do seamless-from-photo algorithms need great improvement yet, the logical things that’s missing is irregular element creation and pure shape/color/pattern based synthesis and detection. One here who’s up to the challenge can try to create a perfect seamless-from-photo pynode and if good enough hardcode it into Blender. As apparently the texture synthesis algorithm as well as about all the others before it fail at perfection on complex images.
What do you think, and should someone code a seamless-from-photo algorithm that adds the things I mentioned?