numberwang on 27 August 2021
There must be more going on under the hood because you can't reconstruct something useful from noise. My guess is that DLSS has access to more data for example a high res wireframe. This would allow to reconstruct stable and accurate shapes from sub pixel details. Nvidia won't tell all their secrets but it is more than just another AI upscaler because these have been around in video and image editing for many years and they never achieved anything comparable.
