There is an article on procedural texture generation by example. I'm trying to implement it, and got lost in the mathematics. http://graphics.cs.kuleuven.be/publications/LVLD10PISTE/ The outline of the algorithm is thus: We start with a noise-like photo. Do a 2D DFT of it: Compute the power band coefficients and reconstruct the image with wavelet noise: Wavelet noise is band limited, which makes it a good base block for such synthesis: That's the theory. In practice, i can't figure out the formula for the coefficients. Formula (E is the input image): Definition of the power band: What does it mean? Should i sum all the magnitudes of the elements in each band? This gives me nonsense if i divide it by the amount of elements, and randomly unscaled value if i don't (that's the above picture, guesstimately scaled). If i sum the real parts of the elements, then i get wildly wrong and out of scale figures. Normalising them also gives me something plausible. And so on, for any way i can think of. I guess i misunderstand something in the notation. Does the integral sign means summing here? Or how should i read it? That's basically the question - what does it mean? Bonus question - anyone heard of this method being used somewhere? A working implementation to peek at would be nice.