I'm curious how the metallic parameter as in the Disney and UE 4 material model could be used.

Following these course notes I implemented the UE 4 style GGX specular model and use metallic like this:

vec3 specular_color = lerp(vec3(1.f, 1.f, 1.f), diffuse_color, metallic);
float f0 = lerp(0.04f, 1.f, metallic);

vec3 specular = specular_color * specular_f(v_dot_h, f0) * specular_d(n_dot_h, a2) * specular_g(n_dot_l, n_dot_v, a2);

vec3 diffuse = (1.f - metallic) * diffuse_color;

return n_dot_l * (diffuse + specular);

From the notes it didn't become quite clear to me what I should do with metallic. It feels strange that non-metallic materials have a specular color of 1, but due to the small f0 it doesn't seem to be a big problem.

Any ideas how this could be improved upon?

Non-metallic materials are achromatic most of a time and thus don't cause the reflections to be colored in any way. In their implementation the more the material is metallic, the more its reflections are colored by the diffuse color. Also the more metallic the material is, the stronger it reflects light in a specular way.

Spec colour and f0 are the same thing. IMO they shouldn't be two different parameters. The Fresnel function will make an RGB f0 term change to white at glancing angles, but if you use an RGB spec colour and a 1D f0 term, you don't get this behaviour.

I implemented it like this:

float3 albedo = ...
float specMask = ...
float metallic = ...
float3 f0 = lerp( specMask.xxx, specMask * albedo, metallic.xxx );
float3 diffuse = albedo * (1-metallic);

Later on I removed the metallic parameter because the artists didn't like having it as basically a second spec-mask parameter, and I replaced it with this:

float metallic = saturate(specMask*2-1);
float nonMetalSpec = saturate(specMask*2)*0.2;
float3 f0 = lerp( nonMetalSpec.xxx, albedo, metallic.xxx );
float3 diffuse = albedo * (1-metallic);

^Basically any spec-mask value below 50% (<186 in sRGB) is a non-metal, with 50%/186 being slightly shinier than diamonds. Above 50% is metallic, with 100% (255 in sRGB) being pure metal. Pure metals use the albedo colour as the spec mask instead, and set the diffuse colour to black. Non-pure metals (187 - 254) use a darkened diffuse term, and a not-fully-coloured specular term.

I used to have a float3 f0 and now I want to fit it back in, following your argumentation.

I notice that you have another parameter called spec mask. Isn't it more confusing to differentiate between roughness and spec mask?

Thanks for the explanation!

After my previous post I remembered the UE 4 course notes to also mention a specular parameter which was later replaced by cavity. I'm having a hard time understand what cavity is used for as it seems similar to roughness.

Picking up your example, lets say we have a highly reflective/metallic material that's been polished. In my case that might be metallic 1 and roughness 0. Now I want to have some small details with less highlights. I could say the roughness is higher here, or the cavity is higher. What is the difference?

In a physically based shader, even if you set f0/spec-mask to 0%, if you view the material at a glancing angle, 1-100% of the light will still be reflected due to Fresnel.
When you have little cavities, lake gaps between plates of armor, this looks weird. In reality, the shadow system should take care of this, but on a low-poly model, that cavity may not even exist (so it can't be shadowed). So, it's like doing a very fina AO bake on your model, or letting you artist hand paint areas where they want to remove the specular reflections.

OK, I never thought about the fresnel still being present on very rough surfaces. I think I can see a use case for the cavity parameter now. Thanks for the explanation!