Dragon (Anisotropic, Surface-like), Fig. 1

Equal sample comparisons of the dragon scene with 512 spps. The noise is stretched by a factor of 100 pointing into the screen. We show that as anisotropy increases toward a heightfield state, the set of admissible paths collapses to a lower-dimensional manifold, causing standard NEE to fail entirely, while our method generates samples directly on this 1D manifold, enables NEE on heightfield GPISes for the first time.

The viewer below uses JERI to compare high-dynamic range OpenEXR files directly in the browser. You can also use the LDR version using PNGs.