Nvidia DLSS 1 and 2 antialiasing discussion *spawn*

[MfA]

Regardless, real time MLPs are very limited in the amount of data they can store. Far too limited for game specific high frequency restoration worth a damn.

MLPs are the dead end which never stops swallowing wasted effort ... computation gives false hope every AI hype cycle, reality keeps smashing it.

 

[iroboto]

Regardless, real time MLPs are very limited in the amount of data they can store. Far too limited for game specific high frequency restoration worth a damn.

MLPs are the dead end which never stops swallowing wasted effort ... computation gives false hope every AI hype cycle, reality keeps smashing it.

i wouldn't agree with that. It's not applicable everywhere but that doesn't mean ti's not applicable. I think what Nvidia has done so far has been promising. they take a 1440p image, AA it, and super resolution to 4K. All of it fitting closely to 60fps on their high end card; I can't see that aspect as being a dead end. The implementation works. There might be other challenges to the pipeline to get more of these types of enhancements out, but that doesn't mean it isn't a problem that can't be solved.

 

[MfA]

I have no idea if it's promising. If the standard for PC games was jittered sampling, reprojection with primitive IDs and coverage data per pixel for analytical SRAA then we could honestly say how much DLSS improves over cheap SRAA ...

PS. I see no reason why it shouldn't be the standard, except for conspiracy theories which the mods get angry about when discussion gets mixed in with anything else, it needs to be quarantined.

 

[DavidGraham]

TomsHardware made a recent comparison of DLSS vs RIS/CAS, in their conclusion, Shadow Of Tomb Raider has a strong DLSS implementation, and DLSS provides a superior IQ to RIS/CAS there, but in Battlefield 5, DLSS is horrible, and it falls way short of any other upscaling solution.

https://www.tomshardware.fr/test-comparatif-visuel-radeon-sharpening-nvidia-dlss/9/

 

[MfA]

Comparing against some bolt on post processing is unfair ... even if you make them temporal the loss of the motion vectors and primitive ID's means you are going to be miles behind compared to a good integrated solution. TAAU and Ubisoft's checkerboarding are the existing competition on the PC side (Insomniac's sample injection seems the most advanced, but is PS4 only). Seeing those in the same game as DLSS is almost certainly not something Meant To Be Played.

DLSS is a tool designed to differentiate and sell hardware, not just relative to AMD but to their own hardware as well. They even had a cheaper method in the pipeline but then apparently decided to bury it in favour of DLSS, Marco Salvi's TSRAA which never saw the light of day. With embedded developers and outright sponsorship it seems unlikely we will have an honest appraisal of DLSS any time soon. All the best work is being done on consoles, because only on consoles there is a lack of perverse incentives :/

 

[Flappy Pannus]

TomsHardware made a recent comparison of DLSS vs RIS/CAS, in their conclusion, Shadow Of Tomb Raider has a strong DLSS implementation, and DLSS provides a superior IQ to RIS/CAS there, but in Battlefield 5, DLSS is horrible, and it falls way short of any other upscaling solution.

https://www.tomshardware.fr/test-comparatif-visuel-radeon-sharpening-nvidia-dlss/9/

That's a great comparison (albeit the image comparison window needs to be larger). If SOTT's DLSS implementation was the norm rather than the exception I think it would actually be a selling point, it just blows away ever other technique there.

 

[MfA]

How can you tell without seeing the actual competition on the same content?

 

[DavidGraham]

DLSS is now featured in F1 2019 along side AMD's FidelityFX (for both upscaling and sharpening). FidelityFX works on NVIDIA GPUs, so it will be interesting to know whether DLSS provides better or worse quality/performance to FX on NVIDIA GPUs.

We have confirmed that FidelityFX Upscaling and FidelityFX Sharpening are both available on Nvidia graphics cards.

https://www.overclock3d.net/news/so...delivers_nvidia_s_dlss_and_amd_s_fidelityfx/1

 

[Deleted member 2197]

This seems to be a pretty good game. Love that late braking on the last lap!

 

[Kaotik]

This seems to be a pretty good game. Love that late braking on the last lap!

Too bad the DLSS is absolute garbage as shown in DavidGrahams link, quite possibly the worst DLSS implementation to date even when we include the first iterations on games where it was improved later

 

[Deleted member 2197]

I agree I've seen better implementations of DLSS. I thought the 5K upscaling was pretty good in the video.

 

[BRiT]

DG's link concluded with:

So, FidelityFX VS DLSS. Is there a winner here? Yes, and no. Yes, because DLSS presented us with too many graphical issues for us to consider it worth using and no because FidelityFX and DLSS aren't directly comparable technologies. Radeon wins by default here, as we can see people using FidelityFX in F1 2019, unlike DLSS.
...
When thinking about DLSS, the first thing that comes to mind is that Nvidia hasn't released any information about DLSS coming to the game. This tells us one of two things, that Nvidia is embarrassed by the graphics DLSS offers here, or that they weren't heavily involved F1 2019's DLSS support. As it stands, DLSS isn't worth using in F1 2019. Just lower your graphics settings in other ways. The performance gains offered by DLSS aren't worth the performance drops they incur. More information about this is on page 2.

Right now, F1 2019 offers some of the worst levels of DLSS support that we have ever seen. Perhaps this will be addressed with a driver update in the future, but as it stands, we recommend that Nvidia RTX users should steer clear from this feature.

 

[iroboto]

Yikes.

 

[MfA]

IMO AMD really should provide devs motion compensated SRAA, even if upscaling with postprocessing can for the moment compare well enough in some cases they can't rely on NVIDIA fucking up indefinitely.

Eventually NVIDIA is going to get the ridiculous GPU load under control and they'll adjust their partner program to make sure fail implementations won't ship. Even ignoring deep learning a proper motion compensated SRAA with surface IDs can have superior AA (see TAAU). Especially with edge masks.

 

[Deleted member 2197]

Yea, apparently CodeMaster's fucked up with the F1 2019 DLSS portion.

Update: Just a quick update on yesterday’s correction to our patch notes for 1.07. We inadvertently released a partially completed integration of Nvidia’s DLSS technology in patch 1.07 which was not supposed to be enabled until a later patch. We are currently working towards enabling this in a completed state for an upcoming patch. Thanks to Nvidia for helping us in tracking down this issue.

https://forums.codemasters.com/topi...box-one-now-live/?tab=comments#comment-451646

 

[Malo]

It took Nvidia to help them realize the DLSS wasn't implemented properly?

 

[Deleted member 13524]

It took Nvidia to help them realize the DLSS wasn't implemented properly?

Probably took nvidia to help figure out the source of the bug.
I find it hard to believe it wasn't nvidia who did all the machine learning for the game in the first place

 

[Deleted member 2197]

If more details are forthcoming it will be interesting. F1 2019 was not on the list of games offering future DLSS support.

 

[jlippo]

Yea, apparently CodeMaster's fucked up with the F1 2019 DLSS portion.

https://forums.codemasters.com/topic/41370-patch-notes-for-107-–-release-info-update-0608-ps4-and-xbox-one-now-live/?tab=comments#comment-451646

I really wonder how long the turn time for seeing results is with DLSS..
Considering the state most of the games start with first implementation, I doubt it's in hours.

 

[Deleted member 2197]

I'm not sure how studios approach DLSS development but do they wait until the game is completed or are they able to process finished parts during the development cycle? AFAIK there is no hierarchy dependency in terms of how images are processed. Gut feeling is DLSS that still something that is being researched and modified as the tech evolves.