Nvidia DLSS 1 and 2 antialiasing discussion *spawn*
- Thread starter DavidGraham
- Start date
-
- Tags
- nvidia
D
Deleted member 2197
Guest
Remnant: From the Ashes to support DLSS
August 19, 2019
August 19, 2019
https://wccftech.com/remnant-from-t...-love-crossplay-targeting-4k-on-ps4-pro-xb1x/
DavidGraham
Veteran
Control features the best DLSS implementation to date, clearly better than regular upscale in most cases.
https://www.overclock3d.net/reviews/gpu_displays/control_rtx_raytracing_pc_analysis/3
https://www.overclock3d.net/reviews/gpu_displays/control_rtx_raytracing_pc_analysis/3
DLSS has a performance cost so this isnt a proper comparison.
?
A proper comparison to what? This thread is about DLSS performance; namely comparative to itself. Is the expectation that DLSS must achieve better image quality with 0 impact to GPU performance to be considered worthy?
When trying to determine if DLSS is better than standard upscaling.
DLSS is being compared to the absolute baseline you can do for upscaling. Why would fairness matter ?
And I would debate whether or not the upscaling uses less power (at least such a large margin that fairness is a discussion). The upscale uses a standard upscale algorithm but AA is still processed at that resolution and therefore a performance impact.
DLSS is trained from aliased images and then constructs the AA and the upsample. As I understand, It uses a modified render path for it to work: one that skips the AA step to produce raw aliased frames for DLSS yo work off.
Its possible that DLSS is being implemented differently here, but in all previous instances it incurred a non trivial performance cost making it substantially slower than standard upscaling from the same resolution. As for why it matters, what use is comparing DLSS IQ to something thats outputting 30% higher fps?
I just assumed that is up to the users to decide. Those who want higher resolution and willing to trade off frame rate for it vs those who are not willing to trade off frame rate for better graphical quality.
The core of the debate around DLSS is how well it can get to it’s target with as little performance impact as possible. Users will naturally choose what they feel is a beneficial trade off for them.
It’s been this way for moving upwards in resolution above 1080p. You’re getting minor improvements at a cost of framerate. Some people can’t see the difference even if its measurably there. PC users get to choose the experience they want.
DavidGraham
Veteran
In the case of control, DLSS runs at the same speed of the regular upscaling solution.
https://www.overclock3d.net/reviews/gpu_displays/control_rtx_raytracing_pc_analysis/7
Well thats much better. If they can get it to be clearly better than standard upscaling across the board it would be a good start.
D
Deleted member 2197
Guest
Looking forward to the Digital Foundry performance review. Dictator answered that DLSS warrants analysis in a separate review from the core RTX review:
https://www.resetera.com/threads/control-pc-performance-thread.137471/page-12#post-23996786
Yeah it looks a lot better than native res in this game and is pretty cheap - definitely something we want to look at further in another video.
Haven't seen it in motion, but in stills it's still blurryYeah it looks a lot better than native res in this game and is pretty cheap - definitely something we want to look at further in another video.
D
Deleted member 2197
Guest
Fortunately we don't play games in stills. I'd be curious to know if this application of DLSS is any different from the existing applications in other DLSS games .... if anything is different regarding implementation on the developers or Nvidia's end.
Last edited by a moderator:
IIIRC quantum break had this issue as well; their temporal solution was very blurry IIRC.Yeah it looks a lot better than native res in this game and is pretty cheap - definitely something we want to look at further in another video.
We have very little information of how devs can implement it, apart now that they can select native resolution.
Can they use it without motion vectors? Without surface IDs? Those will add a few bytes to the g-buffer and with naive implementations perhaps a significant performance penalty. (Though in principle it shouldn't be a big deal.)
Can they use it without motion vectors? Without surface IDs? Those will add a few bytes to the g-buffer and with naive implementations perhaps a significant performance penalty. (Though in principle it shouldn't be a big deal.)
Flappy Pannus
Veteran
Yeah I think this is an important qualifier - native in Control is "native". I haven't seen any mention of an option to disable the upscaling, like Quantum break could do. Granted there's no system alive that could probably run it decently regardless of RTX or not, but it would be nice to see the difference.
So I think this has to be taken into account, at least the extent if Control is being used as an example of DLSS improving. It may very well be, but this may be a one-off case where the fact that Control is inherent blurry due to its reconstruction technique where DLSS just fits perfectly here as it masks the artifacts DLSS can bring. This is not a knock against DLSS, if it works it works here then it works, just saying it may not be an indicator of the technique being significantly improved going forward for other engines.
You certain? DLSS up from 1080p to 4K looks pretty sharp in this game.
This game, unlike, Quantum Break is not using that upscaling from a lower internal resolution by default. So the internal rendering res is the res it says it is.
