Nvidia DLSS 1 and 2 antialiasing discussion *spawn*

[Kaotik]

@Kaotik So if a game supports TAA and DLSS, why is DLSS any less of a "reference" image? You're not modding the game or hacking something into it that wasn't intended.

Edit: Put another way, every single option in a graphics menu is a tradeoff between image quality and performance. Why is DLSS different than any other option? I think anyone that's reviewing GPUs that doesn't include DLSS when it's supported is doing a disservice to their readers. They can offer screenshots and video and let the reader decide for themselves. But to exclude it altogether is absolutely the wrong choice.

Because when benchmarking the testing environment which includes the settings needs to be the same for everyone, obviously? Otherwise why limit yourself to just AA-methods, let's put max settings for all based on what they can do, who cares if it drops RTX-series performance to dust when no-one else is running RT? You see where this is going, right?

@BRiT not a bad idea

 

[trinibwoy]

Because when benchmarking the testing environment which includes the settings needs to be the same for everyone, obviously? Otherwise why limit yourself to just AA-methods, let's put max settings for all based on what they can do, who cares if it drops RTX-series performance to dust when no-one else is running RT? You see where this is going, right?

@BRiT not a bad idea

Clearly independent reviews should cover performance at similar settings. But they should also cover DLSS. The whole point of reviews is to educate consumers after all.

 

[dorf]

It's likely to be game engine issue where motion vectors for birds is not provided.

Maybe, but those ghosty trails can be seen at least in Control, Wolfenstein and Nvidias own NVRTX UE4 branch too. So I'd call it a DLSS issue.

It seems they might be trying to tackle this though (this is the guy who did the DLSS 2.0 GTC presentation responding to someone demonstrating DLSS ghosting in Death Stranding):

https://twitter.com/i/web/status/1284225741115162625

Curiously, something new called "DLSS.Input.Reduce.Ghost" can be found in the very latest DLSS file in Nvidia's UE4 branch (from July) with a hex editor. The previous version from June is missing this. Seems like it could be related to this but could be something else too, who knows.

 

[manux]

Maybe, but those ghosty trails can be seen at least in Control, Wolfenstein and Nvidias own NVRTX UE4 branch too. So I'd call it a DLSS issue.

It seems they might be trying to tackle this though (this is the guy who did the DLSS 2.0 GTC presentation responding to someone demonstrating DLSS ghosting in Death Stranding):

https://twitter.com/i/web/status/1284225741115162625

Curiously, something new called "DLSS.Input.Reduce.Ghost" can be found in the very latest DLSS file in Nvidia's UE4 branch (from July) with a hex editor. The previous version from June is missing this. Seems like it could be related to this but could be something else too, who knows.

DLSS uses multiple frames to pick samples. If the picking and prediction doesn't work perfectly it produces ghosting. Extreme example of this is TAA algorithms that fail really hard in general.

Birds and some particle issues are likely to be motion vector related. If it was inherent to dlss you would see ghosting everywhere similar to birds/TAA. It looks like dlss completely fails at picking samples for birds. Likeliest suspect is motion vectors. Otherwise why is the birds only thing ghosting like that in death stranding?

 

[Scott_Arm]

Because when benchmarking the testing environment which includes the settings needs to be the same for everyone, obviously? Otherwise why limit yourself to just AA-methods, let's put max settings for all based on what they can do, who cares if it drops RTX-series performance to dust when no-one else is running RT? You see where this is going, right?

No, I don't see where it's going. Reviews should benchmark with RT where it's available, and readers can decide whether the performance hit is worth it. Maybe I want to compare a 5700xt vs a 2070 Super to see if the difference in visuals and performance is worth it. This idea of leaving things out because it's not supported on both does not make any sense. Comparing a game running DLSS on one GPU that supports it and TAA on another is exactly what the consumer is going to want to see, otherwise how will they know if it's worth the money for them?

 

[dorf]

DLSS uses multiple frames to pick samples. If the picking and prediction doesn't work perfectly it produces ghosting. Extreme example of this is TAA algorithms that fail really hard in general.

Birds and some particle issues are likely to be motion vector related. If it was inherent to dlss you would see ghosting everywhere similar to birds/TAA. It looks like dlss completely fails at picking samples for birds. Likeliest suspect is motion vectors. Otherwise why is the birds only thing ghosting like that in death stranding?

I'm not disputing this, but I'd just call it a weakness of DLSS when the issue is widespread across most if not all of the game engines in which it is implemented. Probably better to improve DLSS than to "fix" all the game engines.

 

[Malo]

If something like distant birds were not geometry but in fact sprite animations or something similar, I wonder if that could have an effect.

 

[Kaotik]

No, I don't see where it's going. Reviews should benchmark with RT where it's available, and readers can decide whether the performance hit is worth it. Maybe I want to compare a 5700xt vs a 2070 Super to see if the difference in visuals and performance is worth it. This idea of leaving things out because it's not supported on both does not make any sense. Comparing a game running DLSS on one GPU that supports it and TAA on another is exactly what the consumer is going to want to see, otherwise how will they know if it's worth the money for them?

Separate feature specific benchmarks are one thing, general benchmarks are another. I'm talking about the general benchmarks which put each card on a level playing field to see who's fastest.

 

[troyan]

Consumers dont care about sleeping companies. Comparing DLSS with TAA is just fine. It's up to hardware and software companies to provider their own solution.

 

[DegustatoR]

I'm not disputing this, but I'd just call it a weakness of DLSS when the issue is widespread across most if not all of the game engines in which it is implemented. Probably better to improve DLSS than to "fix" all the game engines.

I haven't looked into this specifically but from what I've checked I haven't seen an issue like the one noticed in DS in any other game with DLSS support. What makes you say that it's widespread?

 

[dorf]

What makes you say that it's widespread?

Mostly my own playing and dabbling with Control, Minecraft, Wolfenstein and UE4, but other peoples videos aswell.

Here's someone's example from Wolfenstein:

 

[Benetanegia]

That's 720p DLSS which is actually rendering at 475p... Why? Who? If anything that comparison serves as a "what if the Switch used DLSS" and nothing more.

Almost none of the problems in that video happen at higher resolutions or are vastly vastly reduced... i.e. sparks do definitely exist in normal resolutions, and trees are actually better rendered than on native and TSAA.

 

[dorf]

That's 720p DLSS which is actually rendering at 475p... Why? Who? If anything that comparison serves as a "what if the Switch used DLSS" and nothing more.

Almost none of the problems in that video happen at higher resolutions or are vastly vastly reduced... i.e. sparks do definitely exist in normal resolutions, and trees are actually better rendered than on native and TSAA.

Here's some particles going messy in 1440p, NvRTX UE4 editor. TAA image for reference.

https://imgur.com/a/n9ILKYM

 

[Benetanegia]

Here's some particles going messy in 1440p, NvRTX UE4 editor. TAA image for reference.

https://imgur.com/a/n9ILKYM

Different game/engine tho. I was talking about the other "proof" you provided. As can be seen later in the same video you provided, nearly all the issues mentioned in that 720p portion are gone at higher resolutions. Do they exist? Sure, i.e. at lower resolution in Wolfenstein YB. Are they widespread. When they don't even seem to replicate on same game/engine at different resolutions, I'd say a definitive, nope. Hence it doesn't seem as much of a widespread or inherent issue as you made it out to be, as it is a matter of implementation and resolution combination (or particle size relative to resolution). In other words, it's not a thing to be "fixed" because it's inherent to DLSS. It can still be improved of course. And also Death Stranding really makes the point that motion vector data is in fact required.

 

[pharma]

I think @Dictator did mention that Nvidia was aware of the particle trails in Death Stranding and working on a solution. Though do think they are motion vector related as @manux stated above.

 

[dorf]

Different game/engine tho. I was talking about the other "proof" you provided. As can be seen later in the same video you provided, nearly all the issues mentioned in that 720p portion are gone at higher resolutions. Do they exist? Sure, i.e. at lower resolution in Wolfenstein YB. Are they widespread. When they don't even seem to replicate on same game/engine at different resolutions, I'd say a definitive, nope. Hence it doesn't seem as much of a widespread or inherent issue as you made it out to be, as it is a matter of implementation and resolution combination (or particle size relative to resolution).

In Wolfenstein I can see ghosty trails on 1440p too. Very few though because yes, higher resolution reduces the effect. At core the behaviour is still the same and I see nothing wrong with calling it widespread since it's there to be found on so many of the very few game engines that have DLSS 2.0.

We can just disagree on the definition of widespread though, no big deal.

 

[Benetanegia]

In Wolfenstein I can see ghosty trails on 1440p too. Very few though because yes, higher resolution reduces the effect. At core the behaviour is still the same and I see nothing wrong with calling it widespread since it's there to be found on so many of the very few game engines that have DLSS 2.0.

We can just disagree on the definition of widespread though, no big deal.

Yeah, I think we are disagreeing on the definition of widespread.

 

[BRiT]

widespread, I think refers to occurrence of it happening as opposed to percentage of the screen it impacts?

 

[DavidGraham]

Nobody said DLSS is perfect, but it is the most powerful AA solution (after SSAA), most sharp, stable (flicker free) solution during motion and the fastest one as well compared to TAA/TXAA/FXAA/Temporal Reconstruction/Checkerboarding. These advantages are the main reason why DLSS is held so highly in regards.

 

[xpea]

Nobody said DLSS is perfect, but it is the most powerful AA solution (after SSAA), most sharp, stable (flicker free) solution during motion and the fastest one as well compared to TAA/TXAA/FXAA/Temporal Reconstruction/Checkerboarding. These advantages are the main reason why DLSS is held so highly in regards.

This ^^
The nay sayers are criticizing DLSS when they totally forget to mention that it runs in circles around other solutions that produce far worst result. The truth is that all RTX owners choose DLSS over TAA any day, any time on any game. That's a hard fact and that's why DLSS is such a big deal. It provides better quality than any other available solution AND it boost performance by 30-60%. What not to like ?

EDIT: DLSS 3.0 is coming for Ampere launch (or very soon after), I can confirm that