Nvidia RTX DLSS/Ray Tracing Discussion

[demo]

Originally Posted by jimjobob

DLSS 2.0 is a give and take though. Some parts of the image look better and others look worse. I generally prefer the 2.0 image though because it does a great job with shimmering and usually looks better in motion. I was also a fan of RGSSAA though for the smooth image.

Totally agree on all points

[Exposed]

Originally Posted by Trunks0

https://www.nvidia.com/en-us/geforce...-ai-rendering/



https://en.wikipedia.org/wiki/Deep_l...super_sampling




Bolded. You can dig into DLSS 2.0 and it keeps coming back to working with Temporal Data, which is the exact same information used by Temporal AA and is exactly why nVidia keeps talking about some day getting DLSS to work with any game that has TAA.

You your self pointed out it's correcting the image. Doesn't matter if you refer to that as "correcting" or "reconstructing" the idea is the AI being used to quickly compare and reconstruct based on training with higher quality sample data.

*late edit* Another good read here from AnandTech, NVIDIA Intros DLSS 2.0: Ditches Per-Game Training, Adds Motion Vectors for Better Quality

"It employs new temporal feedback techniques for sharper image details and improved stability from frame to frame."


This is an additional processing component after the fact, the AI is still sampling super high resolution images (much higher than 4k). From that very same link:


A special type of AI network, called a convolutional autoencoder, takes the low resolution current frame, and the high resolution previous frame, to determine on a pixel-by-pixel basis how to generate a higher quality current frame.

During the training process, the output image is compared to an offline rendered, ultra-high quality 16K reference image, and the difference is communicated back into the network so that it can continue to learn and improve its results. This process is repeated tens of thousands of times on the supercomputer until the network reliably outputs high quality, high resolution images.


The bold is all that matters. I can't see this as "correcting TAA" as you're putting in.




Maybe I'm not understanding what you're trying to relay, I'm under the assumption you're stating in DLSS 4k the 1440p image is just TAA upscaled and "adjusted" from there.



The ultra high resolution training has nothing to do with TAA. The temporal component is an additional component required to ensure the next frame, compared from the last frame, is consistent.

[Trunks0]

I can see the confusion, I'm also probably over simplifying.

[Exposed]

I think we're both on the same side of the coin in this argument and don't realize it.

[Trunks0]

Originally Posted by Exposed

The ultra high resolution training has nothing to do with TAA. The temporal component is an additional component required to ensure the next frame, compared from the last frame, is consistent.

Sorta... it's also using that temporal information to build more data to create an image with more detail. Against what it already knows. Think, jitter patterns to build information. It can and does do that to help improve the image.

The AI allows it to "correct" or "Reconstruct" information based on both what it knows things in the image should look like and that is assisted by further knowing whats happening in the image thanks to the temporal data it gets from a given engine's TAA data. Again think jitter'ed data to help the final image come out better.

http://behindthepixels.io/assets/files/DLSS2.0.pdf

Originally Posted by Exposed

I think we're both on the same side of the coin in this argument and don't realize it.

I think so, it's cool stuff

[Exposed]

Originally Posted by Trunks0

Sorta... it's also using that temporal information to build more data to create an image with more detail. Against what it already knows. Think, jitter patterns to build information. It can and does do that to help improve the image.

The AI allows it to "correct" or "Reconstruct" information based on both what it knows things in the image should look like and that is assisted by further knowing whats happening in the image thanks to the temporal data it gets from a given engine's TAA data. Again think jitter'ed data to help the final image come out better.

http://behindthepixels.io/assets/files/DLSS2.0.pdf



I think so, it's cool stuff

I'm still reading that as an additive component process and not "correcting" TAA. Starting from page 57 the DLSS process is described quit well (prior to this page describes the limitations of non neural trained upscaling methods). DLSS 2.0 replaces TAA an anti-aliasing method (because it's equivalent to 64x SSAA) but the temporal component is still dependent (or rather a segment of) the brute force analysis of higher resolution 16k ground truth images. 64xSS, TAA, checkerboarding or any kind of temporal upscaling is not enough to generate a 4k equivalent image on its own (they cover this from page 42). There's still the brute force image comparison analysis that makes up the bulk of the AI process, the temporal segment is a fine tune for the already trained network after the image analysis period is complete.


The only point I'm making is that when running a DLSS 2.0 game in DLSS 4k, it is not a 1440p image with 64x SS applied and then corrected for TAA, which I think is really oversimplifying of what's actually happening.

[Trunks0]

It's just a component. it all comes together to bring the result that is dlss 2.0

[SIrPauly]

I think the temporal component was added to improve compatibility. I would like to see the training improve upon particle like effects.

[SIrPauly]

Since the PDF for Dlss 2.0 was shared, here was Edward Liu of Nvidia discussing the feature. It may bore some but offers some inner workings on the feature:


https://youtu.be/-jrHqUtsM5Y

Personally find this stuff fascinating.

[acroig]

https://wccftech.com/nvidia-official...ytracing-dlss/

[Trunks0]

Originally Posted by SIrPauly

I think the temporal component was added to improve compatibility. I would like to see the training improve upon particle like effects.

The temporal data is one of the biggest changes between dlss 1 and 2. It plays a huge part in it pulling off the current awesome results.

[Ziklitschli]

Originally Posted by Exposed

(because it's equivalent to 64x SSAA).

And yet can't touch a proper x4 with some multisampling heh. Problem is also that it's made of two components, one ruining the other and the other masking the one's bs.

[Ziklitschli]

Originally Posted by Exposed

In Death Stranding, Control, Mechwarrior Mercenaries, Youngblood, Deliver Us the Moon, the results is generally a better, cleaner, more detailed image than native 4k which is in agreement with every website that reviewed DLSS on those games. Not via sharpening either, the reconstructed data (using native high resolution samples) is very similar to native 4k down to pixel detail.

That's bs of the highest order. I actually tried DLSS with MechWarrior some time ago. Anything other than TAA or DLSS (fxaa, off) gave some unbearable dither artifacts or sth. TAA looked like complete cancer it is, and that's what the website compared to. So again, problem is there is AA component and reconstruction component, so saying it's better than native is bs. It will never beat native resolution with a good AA solution, just that smeared crap that games nowadays are, where 1080p with TAA looks like half res or sth, 4k obviously better but still the iq loss doesn't magically disappear, does it.

I read it months ago btw, but they also mentioned the flaws in the article, one about MechWarrior afair. I think you just read what you want to read in those. Anyway, fun fact,
I might actually use it in MW, DSR 4k to 1080p with DLSS quality, that's because all choices are bad in that game. And the fact that the journo draws conclusions from comparision with TAA is just sad.

[Exposed]

You're a fish out of water with your late response. Many others here have tried DLSS 2.0 themselves in newer games and became "DLSS converts". Nothing has changed, only opinions from people who actually tried it on their cards. So still a valid option that does give better than native quality in many cases, not always, but even at it's worst is a much better option than stepping down to native 1440p.

So I don't really understand your negativity about DLSS. Don't like it? Don't run it. I do want to see these cases though where native plus a "good AA" solution is better than DLSS 2.0. SMAA? Try again, DF and HU already delved into this some time ago and found DLSS 2.0 to be a better solution overall, DLSS 2.0 really is the best form of "AA" you can have today, even better than 8x SSAA which will tank your performance otherwise. The fact that DLSS 2.0 can give near indistinguishable IQ from native 4k pixel quality on top of superior AA is a bonus in itself.

But like I said, your arguments are fish out of water. Plenty of people here who dissed DLSS before only to appreciate it, especially in games that benefitted with it on.

[Ziklitschli]

I have no problem with DLSS itself, just statements like "better than native" or "better than 8xSSAA" (I really hope you mean just performance heh) and all that praise of low IQ options. Like digital foundry and all their bs how great and welcome dynamic resolution options on PC are etc etc, yeah sure no dips to 50 but crap going in and out of focus, so great.

Want to see real DLSS, go down to 1080p, turn it on quality and see the perfection, better than native etc and perfect antialiasing in motion heh. Sure 4k it's all less noticable on your probably glossy TV with thick solid glass, or you don't have an eye for it (not a bad thing btw, I envy such people), IDK, but it's all still there.

Do you like dropping settings from ultra to high? The differences are often slight and unnoticeable for most people. Why not med then, motion blur samples, you see that? AO res? Shadow outlines, who looks there? Etc etc. Point is, it all builds up towards image quality, and so does native res against reconstructed one.

There is also, IDK, semantic problem here? All the time you say native, it's native with TAA, even DF said that alll the time afair. As an antialiasing option DLSS is ok, nothing like x64 or even x8 ssaa but is ok. But there's an obvious reconstruction, it's not like native, it might score against materials blurred by TAA but that's it.

So yeah, you like it, run it. Stop preaching it as superior though, it's only superior vs bad TAA or fxaa when it comes to edges. I'm not negative towards it, just told you I'm going to maybe use it in MW, would rather use DSR with smaa/ fxaa but there was that stupid dither (maybe fixed now? have to reinstall) and yes, I'd rather use it over sharpened TAA, plague vs smallpox heh.

[Exposed]

Originally Posted by Ziklitschli

I have no problem with DLSS itself, just statements like "better than native" or "better than 8xSSAA" (I really hope you mean just performance heh) and all that praise of low IQ options. Like digital foundry and all their bs how great and welcome dynamic resolution options on PC are etc etc, yeah sure no dips to 50 but crap going in and out of focus, so great.

Want to see real DLSS, go down to 1080p, turn it on quality and see the perfection, better than native etc and perfect antialiasing in motion heh. Sure 4k it's all less noticable on your probably glossy TV with thick solid glass, or you don't have an eye for it (not a bad thing btw, I envy such people), IDK, but it's all still there.

Do you like dropping settings from ultra to high? The differences are often slight and unnoticeable for most people. Why not med then, motion blur samples, you see that? AO res? Shadow outlines, who looks there? Etc etc. Point is, it all builds up towards image quality, and so does native res against reconstructed one.

There is also, IDK, semantic problem here? All the time you say native, it's native with TAA, even DF said that alll the time afair. As an antialiasing option DLSS is ok, nothing like x64 or even x8 ssaa but is ok. But there's an obvious reconstruction, it's not like native, it might score against materials blurred by TAA but that's it.

So yeah, you like it, run it. Stop preaching it as superior though, it's only superior vs bad TAA or fxaa when it comes to edges. I'm not negative towards it, just told you I'm going to maybe use it in MW, would rather use DSR with smaa/ fxaa but there was that stupid dither (maybe fixed now? have to reinstall) and yes, I'd rather use it over sharpened TAA, plague vs smallpox heh.

If you have a problem with "better than native" then you have a problem with a majority of people who've actually used DLSS 2.0 in multiple games and backed it up with in depth screen by screen analysis.



DLSS 2.0 is clearly better than TAA/SMAA/FXAA and even standard 2x/4x/8x SSAA as an anti-aliasing solution (because it's essentially very high SSAA to begin with).



As a reconstructing solution, DLSS Quality mode is going to be near indistinguishable from native whether 4k, 1440p, or 1080p. There might be some minor artifacts, but cases are these are usually negligable.



And when I say native, I mean with and without AA. Naturally, a non AA native image is going to be sharper / crisper because it's a raw image. But then you'll have eye bleeding artifacts that goes with any non aliased image, so why would you want to compare it to that anyway? You like shimmering? Razor stepped edges?



DLSS 2.0 gives you the benefit of both pixel clarity of native WITH the benefits of AA. No other solution comes close, and this has been demonstrated countless times already.



You want to mention image quality and turning down settings but ironically it is you that has to do these sacrifices if you don't run DLSS 2.0. Especially for ray traced games like Watch Dogs Legion and Cyberpunk. If anything, DLSS 2.0 enhances IQ by allowing you to turn on additional details.



So yes, DLSS 2.0 can be technically "superior" to native resolution with or without AA, and this has been demonstrated numerous times in over a multiple of different tech sites. You also have first hand experiences here right on the forums, even from people who have previously dismissed DLSS as a valid solution.



My suggestion is that if you have a DLSS 2.0 capable card, then maybe play some games other than Mechwarrior Mercenaries. Like the two new games I mentioned or the several other titles previously mentioned. Death Stranding is an example that regardless of performance DLSS 2.0 gives better visual improvement over native. Not my words, the words of many others who played and compared the two.


Oh, and it's not that I appreciate your opinion, it's when you state "Stop preaching it as superior though" when it technically is in many case scenarios, and the benefits far outweigh the drawbacks, and backed by multiple technical analysis from personal and review site coverage, then you have to step back and ask why your stance is a fish out of water. Perhaps you don't have all the information necessary or game data to see the benefits?

[Ziklitschli]

What's 8xSSAA? 4x4? Sparse grid?

And yes, reconstructing from lower res is better than sampling from a higher one because artificial intelligence. Robots man.

[Ziklitschli]

Is it your conclusion btw, the belief in marketing speak or some tech site wrote that? If the latter then link please, so I can never read them again.

Ah and all that fish out of the water thing, I don't care really, and check the initial reactions to mlaa/ fxaa, how it was louded as better than 4xmsaa. The thing was though, all those professional morons were comparing on screenshots, didn't age well did it heh.

[SirBaron]

I don't think anyone ever said FXAA was better than MSAA.

Oh and btw MSAA is dead, and wasn't perfect either.

Shimmering is something MSAA had zero chance of fixing.

[Mangler]

I think he's mixing up TXAA with FXAA.

Or maybe AMD/NVIDIA did try to do some shady mlaa/fxaa marketing in titles with lots of shader aliasing, I honestly don't remember.

[Trunks0]

I also think people get stuck, or hang, on the elements where reviewers and analysis's have pointed to times when DLSS has resolved detail better than native. While forgetting, that they usually also point out points where is fails entirely or removes details. It's not perfect.

I think it's better to look at DLSS for what it is. A compromise that strikes a balance in favour of performance with as little drop in fidelity as they could manage. And it does a damn good job at it.

[acroig]

Originally Posted by Trunks0

I think it's better to look at DLSS for what it is. A compromise that strikes a balance in favour of performance with as little drop in fidelity as they could manage. And it does a damn good job at it.

100% agreed.

[OverclockN']

Originally Posted by Trunks0

I also think people get stuck, or hang, on the elements where reviewers and analysis's have pointed to times when DLSS has resolved detail better than native. While forgetting, that they usually also point out points where is fails entirely or removes details. It's not perfect.

I think it's better to look at DLSS for what it is. A compromise that strikes a balance in favour of performance with as little drop in fidelity as they could manage. And it does a damn good job at it.

Originally Posted by acroig

100% agreed.

Also agreed. I've said it a number of times here on the forum. I look at it as a stopgap for those wanting to play at 4k resolutions, but can't quite get the framerate they want...whether that's 60fps or 120fps. Once we can do that, I see no real need for it.

Logic applies for lower resolutions as well for those that choose not to fork over the cash needed for the master race cards.

[SirBaron]

It will keep getting better with each iteration, to the point where it will always be better than native, whether that's dlss 3 or later.

Native is good for sharpness but terrible for aliasing unless you use taa aka Vaseline.

[Trunks0]

Originally Posted by SirBaron

It will keep getting better with each iteration, to the point where it will always be better than native, whether that's dlss 3 or later.

Native is good for sharpness but terrible for aliasing unless you use taa aka Vaseline.

Well there is always the possibility of doing a version of DLSS that is focused on fidelity instead of speed as well. Maybe an "extreme" mode thats renders at native and just focus's down on killing jaggies.

I mean the core idea of using an AI to handle AA is a good one. It's going to be interesting as we see it expanded on. Because you could potentially use AI in lots of interesting ways with other AA methods as well. Using it with TAA is just the current best method as it's really fast. But could you imagine if it could work with more 3D info than just the current 2D and motion data? Could get some crazy results.

[Ziklitschli]

Yes the tensor cores should be used for proper AA, not some reconstruction + AA mix.

Originally Posted by SirBaron

I don't think anyone ever said FXAA was better than MSAA.

Oh and btw MSAA is dead, and wasn't perfect either.

Shimmering is something MSAA had zero chance of fixing.

It was all over the internet. HardOCP I think posted screenshots comparing msaa and mlaa, can't find it now, maybe deleted and I'd delete it as well, if I was them heh. There are forums mentioning it, I'll browse and post maybe when back from the mountains.

MSAA + OGSSAA or SGSSAA is still king, rarely supported sadly but still happens, thankfully, DCS, ArmA, Total War etc etc. Ofc DLSS is better but hey, 64 > 4, none of us escapes tyranny of numbers.

Yes specular, shadow aliasing is a problem for it (though Wargame series has a separate shadow AA setting) but MSAA + transparency with some light shimmer is 10x better than TAA culled crap and sharpened TAA means noise all over the screen. MSAA times were times when plain 1080p picture looked good, try that now with just TAA and cry.

It's btw devs laziness, MSAA can be implemented, you could mix then with SSAA and SMAA for any leftovers but hey, why not check a TAA tickbox in the engine, people like you will be happy anyway because MSAA dead. I know that firsthand btw, not just guess, and half the time you have to sharpen from outside the game, no in-game slider. Stay happy, DLSS and TAA so much choice.

[Ziklitschli]

Originally Posted by Mangler

I think he's mixing up TXAA with FXAA.

Or maybe AMD/NVIDIA did try to do some shady mlaa/fxaa marketing in titles with lots of shader aliasing, I honestly don't remember.

Nah, I mean MLAA/ FXAA. It really took some time for people to stop lauding it as an MSAA killer, the miracle of progress in AA etc. I don't even mean the usual "4k does not need AA anymore" blind masses, I seem to remember quoting a critical review of FXAA/ MLAA here from a site (AlienBabel tech or sth like that) and have it's credentials questioned because of their negative view of those methods in motion. Anyway this is same crap again, this time end of native instead of MSAA.

DLSS is good btw, for what it is. Should just be put somewhere in medium settings, instead of being treated as a second coming in IQ.

[Exposed]

[LordHawkwind]

Originally Posted by Exposed

Difficult to tell the fidelity difference for me on 1440p. They both looked fine. TBH it's a character walking around no real action to speak of so to me FPS comparisons are meh.

So maybe at 1440p I'd probably get the 4K DLSS frame rate at Ultra RT with no DLSS? Sounds good to me TBH. Is Ultra RT not higher than RT with performance DLSS? Just wondering.

[Exposed]

Originally Posted by LordHawkwind

Difficult to tell the fidelity difference for me on 1440p. They both looked fine. TBH it's a character walking around no real action to speak of so to me FPS comparisons are meh.

So maybe at 1440p I'd probably get the 4K DLSS frame rate at Ultra RT with no DLSS? Sounds good to me TBH. Is Ultra RT not higher than RT with performance DLSS? Just wondering.

Haven't played the game yet so can't speak more about it.

However if you have Xbox gamepass the game is free, so download if you do and post your results. I plan on playing this after Hitmon 3