Shaderball, Fig 7

We display renderings comparing random Fourier features (RFF) and our sparse convolution noise approach for GPISes. Both can render single realizations of stationary GPISes, but our method more closely approximates the ground-truth Gaussian process under few features (27). It also supports non-stationary GPISes with spatially varying length scales and anisotropy, and ensemble light transport obtained by averaging distinct realizations per sample.

The viewer below uses low-dynamic range images. You can also use the HDR version.