Nvidia DLSS 1 and 2 antialiasing discussion *spawn*

[Scott_Arm]

If native has artifacting then there is something inherently wrong with how it's programmed. Thus, DLSS "looking better" than native is only a way of saying that a game is programmed with the arse.

All real-time rendering has artifacts because everything is under-sampled on top of all of the limitations of various rendering techniques, lod scaling etc. When you switch from "native" rendering to DLSS you trade one set of artifacts for another.

Edit: All anti-aliasing solutions like FXAA, SMAA, TAA etc introduce artifacts in the form of blur and the loss of high frequency or high contrast detail, which can then be improved with sharpening which introduces its own artifacts like ringing etc. If you turn AA off you'll get aliasing because you're no longer hiding how under-sampled the image is. You trade artifacts for subjective trade offs in image quality. DLSS is no different, except it comes with a significant performance boost.

 

[Leoneazzurro5]

I remember reading that article when it first released and as with then, I still fail to see any difference in blurriness. If that's the output of DLSS quality then as far as I'm concerned, that's as a good as native.

There are several issues. I.e. The textures on the wooden table or on the floor, or the grates on the floor in the fourth image. Other artifacts may show only when moving. If you don't see differences, congratulations, DLSS is your solution for everything. Someone else could have other opinions, though.

 

[Leoneazzurro5]

All real-time rendering has artifacts because everything is under-sampled on top of all of the limitations of various rendering techniques, lod scaling etc. When you switch from "native" rendering to DLSS you trade one set of artifacts for another.

I know everything in a game is approximated, thanks.The point is you are adding more approximation to everything and calling it better than the original situation.

 

[DavidGraham]

I am questioning people saying "it's better than native"

It is, it's far more temporally stable than native TAA, with less flicker and shimmer, it's also far sharper than native during MOTION, so this is why it's certainly BETTER.

You may pick apart still images all day along for trivial artifacts, but the fact remains that DLSS is sharper and clearer during MOTION than native + TAA. This makes all the difference.

 

[CarstenS]

I remember reading that article when it first released and as with then, I still fail to see any difference in blurriness. If that's the output of DLSS quality then as far as I'm concerned, that's as a good as native.

Did you open the shots in their own tabs and also click into each one in order to zoom to 100%?
The blend-through-white in their gallery function is very irritating for my eye.

The blur (or rather loss of detail) is very obvious in the third row images, on the ground right of the kneeling soldier in the middle of the picture for example.

The palm leafs at the top in the fourth row images otoh look much worse in native resolution than with DLSS enabled.

 

[manux]

One really shouldn't judge dlss2.0 without playing the games. Lot of the errors don't happen all the time and everywhere. Lot of these comparison pictures are result of hunting for errors/differences. Similarly the native or taa has errors on these kinds of comparisons. Also looking at zoomed in stills is very different experience especially in temporal algorithms versus high framerate moving image. Much rather go by either playing the game or look at digital foundry, pcworld, hardware unboxed etc. sites and draw conclusion based on their reviews. Lot of these reviewers recommend turning dlss on in minecraft, control,... as it is the best option considering perf+quality. It's always a compromise and not everyone will prefer same compromise. Some games dlss2.0 doesn't work out well and that is also clearly visible in reviews(f12020 dlss sucks).

I would also tend to trust comments like this from developers

https://twitter.com/x/status/1327713565634035712

 

[Leoneazzurro5]

It is, it's far more temporally stable than native TAA, with less flicker and shimmer, it's also far sharper than native during motion, so this is why it's certainly BETTER.

Really, stop this. Read a lot of comparatives saying the opposite and quit reading only the Nvidia gospel.

 

[DavidGraham]

(f12020 dlss sucks).

F1 2020 uses the old DLSS version.

 

[Leoneazzurro5]

One really shouldn't judge dlss2.0 without playing the games. Lot of the errors don't happen all the time and everywhere. Lot of these comparison pictures are result of hunting for errors/differences. Similarly the native or taa has errors on these kinds of comparisons. Also looking at zoomed in stills is very different experience especially in temporal algorithms versus high framerate moving image. Much rather go by either playing the game or look at digital foundry, pcworld, hardware unboxed etc. sites and draw conclusion based on their reviews. Lot of these reviewers recommend turning dlss on in minecraft, control,... as it is the best option considering perf+quality. It's always a compromise and not everyone will prefer same compromise. Some games dlss2.0 doesn't work out well and that is also clearly visible in reviews(f12020 dlss sucks).

I would also tend to trust comments like this from developers

https://twitter.com/x/status/1327713565634035712

Again, performance with low IQ loss is the word and I won't deny it. Well, I look regularly at these reviews and they ALL say the SAME thing: DLSS is great for being a very good compromise between IQ loss and performance gain. But they also say it's not perfect and you can easily get some artifacts here and there, just look at HW unboxed today's review about 3060 Ti where they EXPLICITELY say that in the conclusion part.

 

[DavidGraham]

Really, stop this. Read a lot of comparatives saying the opposite and quit rading only the Nvidia gospel.

I am not, go watch Digital Foundry comparisons please before spewing the same old concepts that are no longer valid or relevant anymore.

90% of game time is spent in motion, the fact that DLSS is better than TAA+NATIVE in these moments is the decisive factor for anyone seeking stable image.

 

[eastmen]

You've posted DLSS 1.0 vs CAS review which is how AMD proposed to compare them back at 5700 / CAS launch.


Because Fidelity CAS upscaling is just that - upscaling. It doesn't reconstruct anything and thus has a natively worse resulting image quality. There are no comparisons of CAS upscaling with DLSS or TAA SS or CBR because of that. It's worse than any reconstruction.

I hate when i wake up and my posts are in diffrent threads .

My point isn't which is better. Its that DF took 2 weeks to review the card and couldn't be bothered to test its features. If DLSS is better than CAS+ upscaling why not show it ?

The review was extremely devoid of content about the card. Like i also said the last time they talked about FidelityFX CAS+ upscaling was over a year ago. Why not revist it with a new card ? They have done multiple DLSS videos since and they have done videos singing the praises of checkerboard rendering since then also.

 

[DegustatoR]

I remember reading that article when it first released and as with then, I still fail to see any difference in blurriness. If that's the output of DLSS quality then as far as I'm concerned, that's as a good as native.

DLSS in Cold War is completely fine, their "Filmic TAA" blurs the image similarly to DLSS. In this case the resulting quality with and without DLSS is similar enough for the performance benefits to outweigh any minor issues it may incur.

One newer game where DLSS does blur the image noticeably is WD Legion where the default TAA is very sharp. But it comes with a couple of catches too: 1) you can increase DLSS sharpness in settings just as well to reduce the blurring significantly and 2) DLSS AA is in fact miles better than the game's TAA since it provides much better temporal stability overall. So even here saying "but it's blurry" mostly shows that a person didn't actually try it in-game and just looked at static screenshots done at default settings.

F1 2020 uses the old DLSS version.

No, it's using 2.0, just not properly integrated into the game's engine.

 

[Leoneazzurro5]

I am not, go watch Digital Foundry comparisons please before spewing the same old concepts that are no longer valid or relevant anymore.

90% of game time is spent in motion, the fact that DLSS is better than TAA+NATIVE in these moments is the decisive factor for anyone seeking stable image.

You are citing again the ONLY review saying that. Cite all the others saying the opposite.

 

[manux]

I hate when i wake up and my posts are in diffrent threads .

My point isn't which is better. Its that DF took 2 weeks to review the card and couldn't be bothered to test its features. If DLSS is better than CAS+ upscaling why not show it ?

The review was extremely devoid of content about the card. Like i also said the last time they talked about FidelityFX CAS+ upscaling was over a year ago. Why not revist it with a new card ? They have done multiple DLSS videos since and they have done videos singing the praises of checkerboard rendering since then also.

I would think it's because of lack of time. There has been ton of gpu's and cpu's released lately. I fully expect there to be a lot of comparisons once amd gets their newly promised super resolution solution integrated into games.

 

[manux]

You are citing again the ONLY review saying that. Cite all the others saying the opposite.

I just watched hardware unboxed 3060ti review. In that they mentioned they recommend using dlss in watchdogs. So there is at least 2. Can you point out reviews where they conclude don't use dlss in watchdogs?

 

[Leoneazzurro5]

I just watched hardware unboxed 3060ti review. In that they mentioned they recommend using dlss in watchdogs. So there is at least 2. Can you point out reviews where they conclude don't use dlss in watchdogs?

Selective content choice? They said that the option to turn on DLSS is great because the increase in performance with minimal impact to IQ, But they add later "DLSS is super impressive but it's not always flawless sometimes it's suffering from artifacts where it fails to fill the missing information". And again, I am not questioning the value of DLSS as a feature. I am questioning the repeating of the Nvidia mantra as the only absolute truth.

 

[manux]

Selective content choice? They said that the option to turn on DLSS is great because the increase in performance with minimal impact to IQ, But they add later "DLSS is super impressive but it's not always flawless sometimes it's suffering from artifacts where it fails to fill the missing information".

So it's not great to turn on or it's great to turn on? You seem to be sending mixed messages in what you write.

 

[Leoneazzurro5]

So it's not great to turn on or it's great to turn on? You seem to be sending mixed messages in what you write.

No, sorry, my message was clear since first post. DLSS is a good feature that may help achieving good framerates at resolution which are normally not playable with a given graphic card with a low impact on IQ. So if you need it, use it.
What I'm contesting is the fact someone can say it's even better than native solution "because Jensen says so" when it's a (quite advanced) upscaling but as an upscaling it has its share of issues.

 

[Scott_Arm]

I know everything in a game is approximated, thanks.The point is you are adding more approximation to everything and calling it better than the original situation.

What's the original? No AA or TAA? We're talking about PC where user settings are configurable. Some people hate TAA and turn it off. TAA can have significant ghosting and blur depending on the implementation, but textures will always look sharper with no AA than TAA, but then you get aliasing. So what's "the original" and which is "native"?

 

[pjbliverpool]

There are several issues. I.e. The textures on the wooden table or on the floor, or the grates on the floor in the fourth image. Other artifacts may show only when moving. If you don't see differences, congratulations, DLSS is your solution for everything. Someone else could have other opinions, though.

Did you open the shots in their own tabs and also click into each one in order to zoom to 100%?
The blend-through-white in their gallery function is very irritating for my eye.

The blur (or rather loss of detail) is very obvious in the third row images, on the ground right of the kneeling soldier in the middle of the picture for example.

The palm leaves at the top in the fourth row images otoh look much worse in native resolution than with DLSS enabled.

I hadn't because to be honest I couldn't work out how! I was trying to do it from within the DSOG picture viewer rather than from the thumbnail. Anyway, incentivized to put a bit more effort in, I eventually worked it out and was able to flick between both images in full screen mode. I do now see the blurring some refer to in certain parts of some images. However as you note there are also instances where the DLSS image appears to provide more detail. On balance, in a purely academic comparison, I would say native just about edges it out in this game (which DSOG seems to suggest is one of the poorer showings for DLSS 2.0). But in reality and while in motion I highly doubt I would ever notice this, even on a subconscious level. So given it allows for a 25-50% performance uplift depending on the scene, it seems like no brainer to me that you'd turn it on unless you're already hitting 100% of your performance target with everything maxed out.