Bart starts from an 8×8 signal but the DFT has much higher resolution so I’m guessing that he first tiled the Bayer matrix and then took the DFT. This introduces a lot of additional zero entries into the DFT and thus the pixels look more spread out.

]]>Wondering why your fft of the bayer matrix looks different than Bart’s one? These two:

http://momentsingraphics.de/Media/BlueNoise/BayerFFT.png

https://bartwronski.files.wordpress.com/2016/10/2dditheringbayerperiodogram.png

]]>https://thegraphicsblog.com/2018/04/25/controlling-noise-spectrum/

IIRC, the original FFT optimization method (whose patent recently expired I believe) was a lot better, but also somewhat more complicated.

]]>That said, I used it for a pretty basic stochastic optimization problem at work. I’ve not tested how well it’d work for remeshing or any other sort of stress test.

]]>The equivalent of taking a slice out of a 3D blue noise dither array for a point set would be to only consider points with z-coordinates in some small interval and to project these points down to the xy-plane. For a sufficiently small z-interval, that should lead to issues that are very similar to those discussed in the blog post.

]]>http://resources.mpi-inf.mpg.de/ProjectiveBlueNoise/

I used it to generate a progressive sampling mask (like void-cluster), which is a bit of a pain with dart throwing. I think I only managed to generate a relatively small mask, but it was sufficient for my purposes.

]]>