Dragon, Fig. 14

We render the Dragon scene featuring a GPIS with non-stationary length scales. We compare three methods under the Renewal+ memory model: (1) function space sampling from Seyb et al. [2024], (2) our sparse convolution noise method with unidirectional path sampling, and (3) our approach incorporating multiple importance sampling. All three methods, despite their respective approximations, yield renderings that are visually similar. This demonstrates that the approximation errors introduced by both the sparse convolution noise formulation and next-event estimation are minor. All three renderings use 4096 samples per pixel.

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