When I am sampling diffuse BRDF I am using cosine-weighted distribution and when I want to sample specular BRDF I use sampling by GGX distribution. I can decide how much light is reflected and how much is diffused by using Fresnel. That gives me an evaluation formula like this:
Eval = BRDF_specular * F + BRDF_diffuse * (1 - F)
the same principle I can use on PDF. But the problem comes up when I try to use this on sampling. It is obvious that I cannot use this interpolation for the sampled vector. But what I can do is to have some probability that I will sample specular direction and probability for sampling diffuse direction. As sample count goes up it will perform as if it was interpolation between these types of sampling. I can use this for dielectric materials because their Fresnel value is the same for RGB. But because I am using separate values for RGB for Fresnel on metallic material I can have different probabilities (for example, R = 0.8, G = 0.3, B = 0.0). This gives a nice look to metals as you can see in this post. What probability I should choose then? The biggest?? Average? Or is this some different approach to this problem, because I cannot find anywhere how to combine these BRDFs?