Our method is designed to generate materials in the form of SVBRDFs represented as a collection of 2D texture maps. These maps represent a spatially varying Cook-Torrance micro-facet model, using a GGX distribution function, as well as the material mesostructure. In particular, we generate base color $b$, normal $n$, height $h$, roughness $r$ and metalness $m$ properties. The roughness is related to the width of the BRDF specular lobe, where a lower roughness represents a shinier material. Metalness defines which area of the material represents raw metal. We generate both normal and height properties separately, as artists typically include different signals in these maps. We describe the exact model formulations used to compute diffuse $d$ and specular $s$ contributions:
$c_d = b (1 - m)$
$c_s = 0.04 (1 - m) + b \cdot m$
$d = c_d$
$s = \frac{F(c_s, \omega_L, \omega_V) \cdot G(n, r, \omega_L, \omega_V) \cdot D(n, r, \omega_L, \omega_V)}{4}$
$render = (d + s) \cdot IRRADIANCE(n, h, \omega_L)$
Formulations of the different specular components $F$, $G$ and $D$ are the same as Deschaintre et al., and inspired by the Disney BRDF.
Detailed computation of specular contribution:
$\omega_H = \mathrm{Normalize}((\omega_L + \omega_V)/2)$
$D = \frac{\frac{r^2}{(n.\omega_H)^2 (r^4 - 1) + 1}^2}{\pi}$
$G = \frac{1}{(n.\omega_L) (1 - \frac{r^2}{2}) + \frac{r^2}{2}} \frac{1}{(n \cdot \omega_V) (1 - \frac{r^2}{2}) + \frac{r^2}{2}}$
$F = c_s + (1 - c_s) 2^{((-5.55473 (\omega_V \cdot \omega_H)) - 6.98316) (\omega_V \cdot \omega_H)}$