The visual fidelity of a Monte Carlo rendered image depends not only on the magnitude of the pixel estimation error but also on its distribution over the image. To this end, state-of-the-art methods use high-quality stratified sampling patterns, which are randomly scrambled or shifted to decorrelate the individual pixel estimates.

While the white-noise image error distribution produced by random pixel decorrelation is eye-pleasing, it is far from being perceptually optimal. We show that visual fidelity can be significantly improved by instead correlating the pixel estimates in a way that minimizes the low-frequency content in the output noise. Inspired by digital halftoning, our blue-noise dithered sampling can produce substantially more faithful images, especially at low sampling rates.