DavidGraham
Veteran
We spend precious hours fiddling with game settings to increase draw distance for shadows and geometry, and here we have a case of the opposite, arguing to prove the benefit of reducing draw distance!
Nvidia DLSS 1 and 2 antialiasing discussion *spawn*
- Thread starter DavidGraham
- Start date
-
- Tags
- nvidia
I'm a big fan. lol.
I want to see it in person though. I'll know the difference on my 4K screen.
My 2080 runs out of VRAM with 2880p, but I think 2160p is enough to prove that those window frames will become visible as we move towards ground truth.
4K NoAA (imgur doesn't take png over 5mb):
https://ibb.co/g6Yj2mS
1080p & 1440p NoAA & DLSS (fresh shots for accurate comparison)
Zoom in yourselves.
edit: added 1440p TSSAA 8X in the imgur aswell.
Oh yea and 4K TSSAA8X is starting to show those frames faintly now too:
https://ibb.co/ZcH0zrg
4K NoAA (imgur doesn't take png over 5mb):
https://ibb.co/g6Yj2mS
1080p & 1440p NoAA & DLSS (fresh shots for accurate comparison)
Zoom in yourselves.
edit: added 1440p TSSAA 8X in the imgur aswell.
Oh yea and 4K TSSAA8X is starting to show those frames faintly now too:
https://ibb.co/ZcH0zrg
Last edited by a moderator:
neckthrough
Regular
@dorf - thanks for the effort in trying to get closer to some ground-truth images. They are certainly revealing.
BTW -- and this is somewhat orthogonal to DLSS -- looking at these no-AA images reminds me of how awful aliasing is (I know it's more tolerable at 4k but still). And these are static shots - temporal instability in motion would be even worse. It's easy to jump onto the TAA-bashing bandwagon, but man, jaggies suck and TAA does help blunt the pain.
BTW -- and this is somewhat orthogonal to DLSS -- looking at these no-AA images reminds me of how awful aliasing is (I know it's more tolerable at 4k but still). And these are static shots - temporal instability in motion would be even worse. It's easy to jump onto the TAA-bashing bandwagon, but man, jaggies suck and TAA does help blunt the pain.
At what res/ppi does aliasing cease to be a problem ?
neckthrough
Regular
Nyquist rate tells you the sampling frequency from which you can correctly reconstruct the original signal. However, a signal sampled at the Nyquist rate is itself not identical to the original signal - it's still sampled. And so that's not useful in answering the question "at what resolution does aliasing cease to be a problem".
You could ask at what resolution/distance is the angle subtended by each pixel onto the retina so small that the human eye stops perceiving aliasing. There's probably good research on that, but note that Apple's "retina" definition is probably inadequate since it assumes that the source content is antialiased. E.g., text looks crisp on a retina display, but that's because of font antialiasing (e.g., cleartext). Pictures and videos are naturally antialiased because camera sensors accumulate light over an area (instead of a point). In contrast, in computer graphics, each sample represents an infinitesimally small point, the final color value of which is blown up to fill an entire pixel on screen. Without any AA filtering, that s*** is going to be visible to the naked eye at any "retina" setup.
(This was all a layman's take, so I apologize to any experts in the audience who may be able to explain things more precisely).
You could ask at what resolution/distance is the angle subtended by each pixel onto the retina so small that the human eye stops perceiving aliasing. There's probably good research on that, but note that Apple's "retina" definition is probably inadequate since it assumes that the source content is antialiased. E.g., text looks crisp on a retina display, but that's because of font antialiasing (e.g., cleartext). Pictures and videos are naturally antialiased because camera sensors accumulate light over an area (instead of a point). In contrast, in computer graphics, each sample represents an infinitesimally small point, the final color value of which is blown up to fill an entire pixel on screen. Without any AA filtering, that s*** is going to be visible to the naked eye at any "retina" setup.
(This was all a layman's take, so I apologize to any experts in the audience who may be able to explain things more precisely).
I think it’s doing this by default. It’s trained target output is SSAA of the resolution. So if you are targeting say 4K. It should be looking at 4K SSAA as the target output from 1080p.
If you want the Best 1080p output you’re asking for a native DLSS solution with no upscale; so target and trained resolution is same; only that target is using SSAA.
Pretty sure this does work. For example, on a 1080p monitor you could use DSR to render at 4k, then enable DLSS for improved IQ without the severe cost of 4k DSR.
D
Deleted member 2197
Guest
A reference for what the image should be ideally. It's why DLSS was using something like 64x supersampled images as reference for training since it represents the best case for corrected pixels along lines and polygon edges.
Here's a gallery with way too many shots:
Dropped down to 1440x720 to be able up the resolution to 400%. Something breaks when going above 200% resolution which is most noticeable in lines becoming more aliased again. Still, those shots are showing where detail is increasing, like that raytraced shadow on the right.
There's a TAAU 67% resolution shot in there aswell showing how it can improve quality over TAA 67%.
Dropped down to 1440x720 to be able up the resolution to 400%. Something breaks when going above 200% resolution which is most noticeable in lines becoming more aliased again. Still, those shots are showing where detail is increasing, like that raytraced shadow on the right.
There's a TAAU 67% resolution shot in there aswell showing how it can improve quality over TAA 67%.
D
Deleted member 2197
Guest
Which card are you using? I didn't see it mentioned in your other posts.
Edit: Found it ... RTX 2080.
