Oh god not this again.
Adaptive tessellation is a process which tests surfaces for flatness and curvature relative to viewpoint before blindly subdividing the model a fixed number of times (full tessellation). Obviously tessellating a flat surface does nothing to improve the model quality, but the curved part of those pictures is being subdivided the exact same amount of times as the full tessellation model. This is not an example of diminishing returns.
I just threw in it in there for the sake of lolz. Simply put, I just wanted to convey my frustration on why tessellation was previously implemented as an all or nothing solution when this form shows the biggest gains come from selective tessellation and beyond a certain point the performance hit is not commensurate with the output.
I hope ND can pull it off at least with characters; recent showing of Snake's fingers left me cold.