yep but that's a minefield.. any optimization or trick or adhoc implementation arnold is doing can broke everything.. it's good because so I learn some internals.. ie fixed everything on beauty pass without performance drop but .. for example with aovs on a first try I can only trace binary/clipped opacity ..


ie the borders should be semi-opaque as in the beauty
another problem is that on specular aovs it breaks somehow sampling .. if you look close you can see the quad holding the leaves producing bad sampling especially on the upper right border .. where it should be fully transparent.
