So I'm adding spherical area lights to my application, and comparing my results with mitsuba, I am getting some differences (left is my approach, right is mitsuba - a pathtraced reference):
What I am mainly noticing is that:
The specular is way too bright for roughness values over 0.5
The specular shape gets "flat" (Left-mine, Right-mitsuba, noise caused by low sample count):
I am using the closest point approximation and the normalization term proposed in Real Shading in UE4, that is:
(radius / (distance * 2.0) + alpha)²
I believe issue 1) could be an issue with my normalization term, however I think issue 2) might probably be an issue with the closest point approximation?
Is this kind of inaccuracy expected with the approximations I use, and if so, are there better approximations?