I read some more about it and it looks like developers have a lot of granular control. Not just % applied but options per object type. So they can max it out for faces, 50% for water, 25% for foliage, etc.

There are some legitimately awesome use cases for this especially if they let developers train their own models. I didn't play Death Stranding but I know they've got detailed face scans of Norman Reedus...imagine if the Norman filter got applied to his character in-game.