Hey have you gotten motion vectors to behave properly your version? Without MVs any small movement causes lighting and details to smear around. I have them working 70% of the time for my plugin but 30% of the time some funky scene setup causes them to break in fun ways.
Yo! : )
Nope I'm not talking about smearing... I saw that effect when testing it on VAM2, and indeed that's a problem about motion vectors. In this case, I'm talking about actual lighting computation and AO interpretation. A good test case is just moving towards a character at a steady pace. You see the AO adapt and change over several frames. It's clearly a limitation of the diffusion system, as it probably can't be adequate/accurate at every distance.
I'll post a set of situations you often see if you go further than just slapping the thing on for a quite standard scene.
The conclusion I got to is pretty much "that works well under normal lighting circumstances, anything quite artistically driven immediately creates issues".
Which technically, feels logical. Most diffusion models on people was trained under very standard lighting (and I'm not even accounting for the fact that models have insane biases)... which means that anything non-standard will yield results that you can't control much no matter what Huang wants to believe.
And the settings to control intensity is just some sort of merge of both image pre and post diffusion... at least, its how it feels, which makes the result somewhat strange. (when they could probably have made several different models yielding a less heavy diffusion result).
That's probably a fun tool for a porn game like VAM (at least for people who like images that heavily lean on slop)... for production ready games, that thing is gonna create all sorts of issues sadly. And I highly doubt studios are gonna Q/A the shit out of their games to ensure every single important spot of their games are looking good with and without NR.