Scouting ninja - in your first post you've said that "gloss" can be used as a name for two of the maps (intensity and size) -- this is why using roughness/smoothness/etc is now common.

You've bolded the one under size though -- this is the same as roughness, it defines the size/shape of the specular highlight, while the other map (intensity) controls how visible the highlight is. The intensity map can be RGB or monochrome, or both (which will be multiplied together in the shader).

In traditional ad-hoc shaders, artists needed to tweak the two maps in conjunction to get good results -- e.g. increasing intensity while reducing size.
With PBR, this is no longer the case. We're now using real physics (regardless of whether were raytracing or not, the BRDF is the same) - so the material inputs / maps that we provide are now physical values. A *normalized* specular BRDF means that the intensity is automatically boosted/dulled by the physically correct amount based on the roughness/size value. The intensity is also modified according to Fresnel's law; the value painted in your specular intensity map is the fresnel reflectance at normal incidence, Fresnel's equations allow us to use that value to find the correct specular intensity for other incidence angles.

An infinitely smooth surface will reflect 100% of the incoming light out of one single point, a Lambertian-rough surface will reflect 1/pi of the incoming light from each point - that amount is then multiplied by the 'intensity' value, which results in small highlights being brighter than large highlights, even if both materials have the same specifed 'specular intensity' value.

With a PBR renderer, the spec maps are not just values to fool the viewer, they are real, measurable physical values being plugged into a simulation. Also, a lot of the same BRDFs are being used in real-time engines now, as are being used in offline ray-tracers for VFX. While VFX has the liberty of Ray-tracing everything, they still often use IBL for "ambient specular", which still gives real reflections, very quickly, but only for distant objects -- real time systems now use the same IBL type systems as film too!

The correct specular intensity value for a material can be calulated if it's IOR is measured -- for most non-metal objects, this value is about 3/100 -- this value does not change depending on whether the surface is smooth/glossy/rough, it only depends on the raw material that the surface is built out of.

Roughness/smoothness/size/etc is a statistical value that gets plugged into a probability distribution to simulate a highly detailed normal-map that's smaller than a pixel. If you cared to sculpt/scan your surfaces at the micro-meter scale, you could automatically select correct roughness values based on the distribution of the normals within that micro-surface.

So basically - your advice is correct for traditional rendering systems, but not PBR systems.

Under a traditional (non-normalized) BRDF:

-- smooth plastic has low roughness / high gloss, and has high intensity.

-- rough plastic has high rougness / low glass, and low intensity.

Under a PBR BRDF though, they're both the same material, and they're both a common dielectric, so they both have the same intensity of about 3% (AKA Fresnel reflectance at incidence zero) and only the roughness/gloss value is different.

Likewise, smooth plastic and smooth chrome share the same roughness/gloss value, but have completely different 'intensity' values (~3% vs 90+%).

[edit] Here's another link dump - not so great for implementation, but just talking about PBR:

http://www.marmoset.co/toolbag/learn/pbr-practice

https://www.fxguide.com/featured/the-state-of-rendering/

http://www.fxguide.com/featured/the-state-of-rendering-part-2/

http://www.fxguide.com/featured/game-environments-parta-remember-me-rendering/

Yes.

Though even Phong or Blinn-phong (the typical specular formulas) both stored intensity and roughness values in the g-buffer too (often called spec-mask / just "specular" and spec-power / spec-exponent).

But from what I've undersood the specular power is actually the inverse of roughness, since a roughness value of 0 would correspond to a very high gloss value (over 100 at least or something?). You're right that the change in the G-buffer isn't much of a change though.

Yeah. But in practice, roughness is often actually stored in a "smoothness" map, where 0=rough and 255=smooth... but we still call the variable roughness anyway

With specular power, it's a value that goes from 0 to infinity -- or generally, usually from 1 to a large number, maybe 2048 or 8192...

Often to decode spec-power from a texture (which only stores a 0-1 range), you'll use something like:

float power = sqrt( tex2D(...) ) * 8192 + 1; //or

float power = pow( 2, tex2D(...) * 12 + 1 );

Each engine will decide on a different mapping between their textures and how to decode them into spec-power values...

Different BRDFs will use different ranges for roughness. In some of them, roughness is in the 0-1 range, where 0=rough and 1=smooth -- in that case, you can just use the texture as is. In others 0=smooth and 1=rough -- in that case you can just use 1-tex2D(...). With the GGX BRDF that I'm using, I'm currently using 0=smooth, 1=rough and 4=extremely-rough, so I use pow(tex2D(...),4.4)*4.

The different curves - pow(x, y) or sqrt(x) (which is just pow(x, 1/2.0)) are there to skew the middle grey levels in the texture either towards 1 or 0.

e.g. you might want spec-power values from 1-8000. However, spec values of 1.5, 2, 2.5 and 3 give extremely different appearances, while spec values of 7800, 7900 and 8000 all look pretty much identical. So, in this case, you want to skew distribution of your middle grey values down towards zero, so you've got a more even distribution of different appearances when the artists try to paint their maps.