Wed May 15, 2019 11:42 am
From a bit of futzing around and comparing results with explicit transforms in Nuke it seems that in v16 layer mixer does its blending not in project working space, which for ACES project is probably cc or cct, but linearizes the inputs first and transforms to ACES2065-1. This looks like the main cause for this difference. Why it was changed and whether it is documented... dunno.
But wait, there's more! Blur operation that you use as input into layer mixer is done in working space. So there is a lot going on in this little freq separation thingy.
To replicate v16 behavior in Nuke, something like this should be done:
- input is transformed to working space, for example ACEScct, with IDT;
- branch out the blur and apply blur operation;
- on both branches, apply ACEScct > ACES2065-1 transform;
- do subtraction (minus) operation;
- convert back to working space ACEScct;
- apply view transform
My guess why the change is that somewhere along the way someone discovered that blending ops should be done on linearized values (which is ok), but as everything else is so strongly built on top of using some aces log curve as basis for grade ops, that part was left untouched. So now there are implicit colorspace conversions happening in some selected parts of node tree with no indication about it.
I do stuff
