Soldato
Most of the effect you see on DLSS is the edge detection and sharpening of edges. This why it looks better and is compared to TAA/FXAA which are generally softer than a native image, especially the edges. It's the same effect when people look at pictures from a phone/compact when compared to native ILC output and the former looks "better". It's the post processing you see in play and Nvidia knows this very well.It's the same reason why digital used to look better than even higher resolution film at the start. That "sharpness" which is localised contrast enhancement is very perceivable by human vision,even if there is actually less detail. It's why certain things like USM are very useful tricks in photography. There are other computer tricks which also work very well with out vision.For instance almost all digital sensors are black and white with a RGBG colour filter on top - 2:1:1 green to blue to red. So each pixel you see is one colour and the other two are predicted. The green bias is there because human vision is biased towards green. To reduce colour artefacts most sensors have a filter over the top which slightly blurs the image on purpose.
DLSS is itself normal upscaling especially in the latest version,but what is clever is that I suspect it is being trained on a per scene basis. So what you probably have is it switching to scene optimised upscaling algorithms.
The first version of DLSS was actually meant to be tailored per game but from my experience of seeing how machine learning works I did tell others here it would take too long to have support in games, because of the amount of training required.
Moving to a more generalised way was probably the better way forward.
DLSS is itself normal upscaling especially in the latest version,but what is clever is that I suspect it is being trained on a per scene basis. So what you probably have is it switching to scene optimised upscaling algorithms.
The first version of DLSS was actually meant to be tailored per game but from my experience of seeing how machine learning works I did tell others here it would take too long to have support in games, because of the amount of training required.
Moving to a more generalised way was probably the better way forward.
Last edited: