Constant-time energy-normalization for the Phong specular BRDFs

Abstract

The Phong and Modified Phong specular BRDFs, although of limited physical basis, are nevertheless some of the simplest BRDFs exhibiting glossy and specular qualities to understand and to implement, making them useful for validation and teaching. Unfortunately, although it is well-known how to make these BRDFs conserve energy (that is, never gain energy), making them energy-normalized (that is, never lose nor gain energy) is far more difficult. Lesser-known algorithms exist, but require the specular exponent n to be integer-valued, and have O(n) runtime cost. We express these algorithms as mathematical formulae and generalize to the real-valued specular exponent case. We then simplify and optimize to finally attain an algorithm that is O(1). Energy normalization makes the Phong BRDFs more physically plausible and therefore both more practically and theoretically useful—and our improvements allow for this energy normalization to be done efficiently and without arbitrary limitations.

Appendices

A: Derivation of energy normalization criterion

The following informal proof demonstrates the well-known BRDF normalization criterion. We start with 1:

$$\begin{aligned} 1&{=} \frac{L_o}{L_o} {=} \frac{1}{L_o} \int _{S^2} d L_o {=} \int _{S^2} \frac{d L_o}{L_o ~ (\varvec{N}\bullet \varvec{\upomega }_o) ~ d \varvec{\upomega }_o} ~ (\varvec{N}\bullet \varvec{\upomega }_o) ~ d \varvec{\upomega }_o \end{aligned}$$

Then, since the incoming irradiance $E_e = d \Phi _i/d A$ must match the outgoing radiant exitance $M_e = d \Phi _o/d A$:

$$\begin{aligned} 1&= \int _{S^2} \frac{d L_o}{d M_e} ~ (\varvec{N}\bullet \varvec{\upomega }_o) ~ d \varvec{\upomega }_o = \int _{S^2} \frac{d L_o}{d E_e} ~ (\varvec{N}\bullet \varvec{\upomega }_o) ~ d \varvec{\upomega }_o =\\&= \int _{S^2} f_s(\varvec{\upomega }_i\rightarrow \varvec{x}\rightarrow \varvec{\upomega }_o) ~ (\varvec{N}\bullet \varvec{\upomega }_o) ~ d \varvec{\upomega }_o \end{aligned}$$

B: Derivation of energy conservation

Unlike energy normalization, which is the subject of this paper, the derivation of energy conservation terms for the Phong BRDFs is widely known.

Consider the Modified Phong specular component BRDF. In Eq. 5, first assume that ${\varvec{\upomega }_i = \varvec{N}}$:

$$\begin{aligned} I_M&= \int _\Omega \text {max}\left( 0, \varvec{R}(\varvec{\upomega }_o) \bullet {\varvec{N}} \right) ^n \left( \varvec{N}\bullet \varvec{\upomega }_o\right) ~ d \varvec{\upomega }_o \end{aligned}$$

Then, since the dot product of the reflected vector must be nonnegative and the same as that of the original vector:

$$\begin{aligned} I_M&= \int _\Omega {\left( \varvec{\upomega }_o\right. } \bullet {\varvec{N}} {\left. \right) }^n \left( \varvec{N}\bullet \varvec{\upomega }_o\right) ~ d \varvec{\upomega }_o\\&= \int _\Omega \left( \varvec{N}\bullet \varvec{\upomega }_o\right) ^{n+1} d \varvec{\upomega }_o\\&= \int _0^{2\pi } \int _0^{\pi /2} \text {cos}^{n+1}(\phi ) ~ \text {sin}(\phi ) ~ d\phi ~ d\theta \\&= \int _0^{2\pi } \frac{1}{n+2} ~ d\theta \\&= \frac{2\pi }{n+2} \end{aligned}$$

The computation of $I_P$ with Eq. 4 is analogous.

C: Example code

Although the normalization terms given by Eqs. 7 and 8 may look intimidating, we provide a listing of the key source code, with simple optimizations applied, to demonstrate that our approach can be implemented in only a few lines:

The incomplete beta function is the only omission from the above, since it is not provided by most implementations of the C/C++ standard library. However, there is no shortage of available implementations to choose from. For example, the function from Boost [3] will be available to many users; the additional implementation would look like:

In our experiments, we chose to adapt the implementation of [11], because it seems to be based on a newer method. Such choices also in principle affect the epsilon at which the incomplete beta term must be replaced with its limit. However, empirically, our implementation was stable down to $n \approx 5 \times 10^{-37}$, so a very small epsilon, as-above, is still adequately conservative.