Mitchell’s best-candidate algorithm generates a new random sample by creating k candidate samples and picking the best of k. Here the “best” sample is defined as the sample that is farthest away from previous samples. The algorithm approximates Poisson-disc sampling, producing a much more natural appearance (better blue noise spectral characteristics) than uniform random sampling.
See also the white-on-black and Voronoi variations of this example.