Is Toksvig better than VMF,LEAN or CLEAN algorithm ?

All of these algo give different result, what is the one who is more correct ?

It's funny that no-precomputed need to convert to specPower and then back to roughness :

normalTS = NormalMap.Sample(AnisoSampler, uv).xyz * 2.0f - 1.0f;
float normalMapLen = length(normalTS);
float s = RoughnessToSpecPower(roughness);
float ft = normalMapLen / lerp(s, 1.0f, normalMapLen);
ft = max(ft, 0.01f);
roughness = SpecPowerToRoughness(ft * s);

Here the precomputed :

roughness = RoughnessMap.Sample(LinearSampler, uv).y;
roughness *= roughness;

I have compare both version, the precomputed give a result who is more near of the disabled in the MJP sample.

That surely end to precision problem I guess with all these conversion.

I don't understand why the precomputed give different result.

Is super sampling a must have ?

LEAN mapping looks the best, and takes up the most overhead. There was a great demo where you could switch between LEAN, CLEAN, and maybe Toksvig too, but my google skills are failing me.

MJP demo does it good about the switch, but not really easy to understand all that correctly.