How do you like DLSS?

[DCoffey]

On 1/10/2024 at 4:07 PM, Nedum said:

It has !!nothing to do with "the setup". If one thinks DLSS adds no blur, the person has bad eyes. DLSS will !!always!! add blur. That's how DLSS works to gain more FPS.

I bet $1.000 I can show every guy who tries to tell there is no difference, that there is a huge difference in the image quality and that there will be always blur.

It has nothing to do with VR, 2D, the setup or the used hardware.

The technic behind DLSS can never be as good as the original. That DLSS could be as good as the original DLSS would have to be a time machine so DLSS knows !!before!! the image would be rendered what will happen in the future and that, for sure, will never ever happen.

If one tells us there is no noticeable difference between DLSS and the original picture, there could be two reasons for:  1. the person can't see it, because he needs glasses or new ones, 2. the person do not want to see it.

DLSS will always add blur to the image. Now and in the future.

 

Well aren’t you just negative. You tried it on your system with your settings and it didn’t work for you. For me DLAA on my 4 series card was a huge step up in getting rid of sparkling, jaggies and a nice increase in fps for me in vr. 

[Fisherman82]

17 hours ago, DCoffey said:

Well aren’t you just negative. You tried it on your system with your settings and it didn’t work for you. For me DLAA on my 4 series card was a huge step up in getting rid of sparkling, jaggies and a nice increase in fps for me in vr. 

Same for me with a 3080. When the 2.9 update came all shimmering that was massive before disapeared. Instead came much better FPS.

And Im running DLSS instead of DLAA. Cant see a reason for using DLAA because it gives me a bit less FPS and I cant see all the DLSS problems some are talking about. To me 2.9 was a wild improvement.

But one thing to note, Im not running turbo mode in Open XR, perhaps the guy who mentioned that is on to something.

[Qcumber]

2 hours ago, Fisherman82 said:

But one thing to note, Im not running turbo mode in Open XR, perhaps the guy who mentioned that is on to something.

I tried toggling it on and off during a mission and saw no difference. 

 

2 hours ago, Fisherman82 said:

And Im running DLSS instead of DLAA. Cant see a reason for using DLAA because it gives me a bit less FPS and I cant see all the DLSS problems some are talking about. To me 2.9 was a wild improvement.

DLAA looks a bit better to me. Not sharpness as such, but an overall clarity of image. It introduces more contrast which makes buildings look a unnatural. Reducing contrast with OXRTK helps a bit with this. 2.9 is a really big improvement overall for me. It's great. 

[Nedum]

On 1/14/2024 at 5:37 PM, DCoffey said:

Well aren’t you just negative. You tried it on your system with your settings and it didn’t work for you. For me DLAA on my 4 series card was a huge step up in getting rid of sparkling, jaggies and a nice increase in fps for me in vr. 

 

It has nothing to do with the System. The technic dictates what we all will get and can see!         

 

But who I am to forbid one to believe in their "DLSS-God"?

If one "can't" see the obvious, I am the last person who wants to fight with a believer. I know it better, for sure, and that's enough for me.

If one wants to learn more about DLSS and DLAA, please ask, but I will never tell one what one want to hear, I will tell one the truth, nothing more and nothing less.

The truth is: DLSS is nothing for fast path games and DLSS will !!never!! reach the image quality of the original image. That's a rule of the used technic behind DLSS. That's the truth. And as I said before, that has nothing to do with the system one is using.

[twistking]

4 hours ago, Nedum said:

 

It has nothing to do with the System. The technic dictates what we all will get and can see!         

 

But who I am to forbid one to believe in their "DLSS-God"?

If one "can't" see the obvious, I am the last person who wants to fight with a believer. I know it better, for sure, and that's enough for me.

If one wants to learn more about DLSS and DLAA, please ask, but I will never tell one what one want to hear, I will tell one the truth, nothing more and nothing less.

The truth is: DLSS is nothing for fast path games and DLSS will !!never!! reach the image quality of the original image. That's a rule of the used technic behind DLSS. That's the truth. And as I said before, that has nothing to do with the system one is using.

If you have some sort of god complex, it must be truly difficult to endure the opinion of heretics.

[Art-J]

6 hours ago, Nedum said:

If one "can't" see the obvious, I am the last person who wants to fight with a believer. I know it better, for sure, and that's enough for me.

Then stop acting like "believer" yourself. Nobody in this thread said DLSS gives better quality than raw rendering. We all know it comes at a price of some quality. But acceptable or "noticeable" quality reduction from using it is a personal preference and different for everyone, just like acceptable minimum framerate, or acceptable level of stuttering. You seem to be blind to that fact, though.

 

[twistking]

11 minutes ago, Art-J said:

[...] Nobody in this thread said DLSS gives better quality than raw rendering. We all know it comes at a price of some quality. [...]

I don't have an opinion on wether that's true with the current state of the technology, but in theory (!) a technique like DLSS could archive higher quality than raw rendering even if the raw image is rendered with MSAA. A technique that uses subpixel shifting with motion vectors to combine shifted images in a temporally stable way (what DLSS does, or tries to do) can archive a significantly higher effective resolution. People tend to not understand this: Techniques like DLSS accumulate image information over subsequent frames allowing it in theory to reconstruct an image with higher detail than "ground truth". The ideal case would be a static scene with a high frame rate. In the typical case (motion scene, lower framerate) the quality of the reconstruction is highly affected by the success of the temporal matching. In practice this is what leads to the perceived blurriness or -in the worst case- ghosting.

Interestingly there have been many related techniques used to enhance video capture in the real world. For example subpixel shifting for ultra high resolution still photography.

[Dragon1-1]

Note that, no matter what happens, you physically can't display any more detail than you have pixels on the screen. So, a regular with supersampling (not MSAA) will always produce better results. Temporal integration can substitute for a static or slow moving image, but the more motion there is, the less capable it becomes. Note that it's different from photos and videos, where generating a larger image than the sensor's physical resolution is not a problem. DLSS and related techniques are ultimately band aids, they don't actually create new information in the image, they just try to fill the gaps by guessing.

I think ghosting could be alleviated to some extent by tweaking things internally. I don't know how much customization is possible with DLSS algorithm, but it's got to have some knobs that the programmers can fiddle with. It'll remain, but it should be possible to reduce it. For instance, shorten the time interval for temporal integration.

[twistking]

17 hours ago, Dragon1-1 said:

[...] So, a regular [renderpass] with supersampling (not MSAA) will always produce better results. Temporal integration can substitute for a static or slow moving image, but the more motion there is, the less capable it becomes. [...]

Yes, this should be ridiculously obvious: A highly supersampled frame can be seen as ground truth and -in practice- Nvidia will likely have used highly supersampled footage for their deep learning training data. High sample supersampling is the idealistic gold standard for rasterized graphics, but it is simply not economical.

17 hours ago, Dragon1-1 said:

[...] Note that it's different from photos and videos, where generating a larger image than the sensor's physical resolution is not a problem. [...]

It's obviously different, because it was only an analogy. Digital photography also deals with the problem of "rasterization" and does so in different ways. Subpixel shifting in DP fulfills the same goal as it does in computer graphics rasterization: Retrieving a higher amount of spatial detail frequency from a theoretical "ground truth". Again it was meant as an analogy only to get people thinking about subpixel shifting.

17 hours ago, Dragon1-1 said:

[...] DLSS and related techniques are ultimately band aids, they don't actually create new information in the image, they just try to fill the gaps by guessing. [....]

Maybe we are discussing semantics now, but i think this statement is incorrect or at least misleading. Even without the "deep leraning" aspect, i would not call the upscaling a "guess". Technically there is no "guessing" (or RNG) going on, so i feel that you use the term simply in a derogatory way. MSAA and texture filtering could also be described as band-aids, that only do mere guessing to "fill in the gaps".

17 hours ago, Dragon1-1 said:

I don't know how much customization is possible with DLSS algorithm, but it's got to have some knobs that the programmers can fiddle with. It'll remain, but it should be possible to reduce it. For instance, shorten the time interval for temporal integration.

I'd assume that developers can most importantly work on cleaning up or adding motion vector data, or maybe even reserve certain assets from DLSS altogether, when good vectors can't be generated.  I'd also assume that improving frametime health could help with temporal integration and reduce ghosting, by allowing the algorithm to make more precise motion predictions.

Edited January 18, 2024 by twistking

[Dragon1-1]

1 hour ago, twistking said:

Maybe we are discussing semantics now, but i think this statement is wrong or at least strongly misleading. Even without the "deep leraning" aspect, i would not call the upscaling a "guess". Technically there is no "guessing" (or RNG) going on, so i feel that you use the term simply in a derogatory way. MSAA and texture filtering could also be described as band-aids, that only do mere guessing to "fill in the gaps".

MSAA and texture filtering don't attempt to generate new pixels, just more accurately place the existing ones. They're inferior to SSAA, so they're band aids in that sense, but it's not the same thing as upscaling. DLAA falls into the same category. Quite frankly, they're all necessary only because our way of turning vectors into pixels on a screen isn't perfect (not sure if it could be, even with a paradigm change, my knowledge is mostly from the imaging side of things).

"Guessing" doesn't refer to RNG, in fact, a guess doesn't need to be random at all. It's just term for making up information you don't have. The very basis of upscaling algorithms is to actually render fewer pixels than the display can show, then guess those that were not rendered. That's literally what DLSS does. The actual information carried by an image is contained within "real" pixels it consists of. An upscaling algorithm can't actually add information, but it can do is present what's there in a more visually pleasing way. However, if it guesses wrong (as it is liable to when there's a lot of motion), those added pixels will be incorrect. Deep learning just allows it to guess much better, but not in a perfect way. You're always trying to predict information that is simply not there. 

Oh, and you don't really need extreme, "gold standard" supersampling, SSAA had been a thing for a while. It's performance intensive, but usable, and gives better results than DLSS if you can run it.

[RealDCSpilot]

On 1/10/2024 at 5:34 PM, mobile83 said:

My question: how can you like DLSS?

Because it's the best VR anti aliasing option since... years? The super sampling from 2496x2496 -> 3744x3744 per eye works really well for me. Thanks ED!

[rfxcasey]

It's great. I can run at 120FPS+ with full detail without DLSS on, but it reduces my GPU temp by about 10 degrees with no negative consequences.

Edited January 18, 2024 by rfxcasey

[twistking]

On 1/18/2024 at 6:56 AM, Dragon1-1 said:

MSAA and texture filtering don't attempt to generate new pixels, just more accurately place the existing ones. They're inferior to SSAA, so they're band aids in that sense, but it's not the same thing as upscaling. DLAA falls into the same category. Quite frankly, they're all necessary only because our way of turning vectors into pixels on a screen isn't perfect (not sure if it could be, even with a paradigm change, my knowledge is mostly from the imaging side of things).

"Guessing" doesn't refer to RNG, in fact, a guess doesn't need to be random at all. It's just term for making up information you don't have. The very basis of upscaling algorithms is to actually render fewer pixels than the display can show, then guess those that were not rendered. That's literally what DLSS does. The actual information carried by an image is contained within "real" pixels it consists of. An upscaling algorithm can't actually add information, but it can do is present what's there in a more visually pleasing way. However, if it guesses wrong (as it is liable to when there's a lot of motion), those added pixels will be incorrect. Deep learning just allows it to guess much better, but not in a perfect way. You're always trying to predict information that is simply not there. 

Oh, and you don't really need extreme, "gold standard" supersampling, SSAA had been a thing for a while. It's performance intensive, but usable, and gives better results than DLSS if you can run it.

i see, that we both more or less agree and that the argument is about semantics mostly: i feel that my words are a better description of what's happening, but that's a futile discussion. even more so because english is not my first language

one technical point though: isn't SSAA just good ol' supersampling? i'm pretty sure that internally it's the same. so SSAA on "x2" should just render with double or (quadruple?) the pixels and then downscale. maybe 2d UI elements are handled differently in the pipeline, but that should not have a big impact.

[Aluminum Donkey]

I think that DLSS is awesome! Just not for flight sims. In anything else, it makes my 3070 perform like a 3090 Ti

In DCS though the little "crumbly and wiggly" effects are noticeable even on the Quality setting. In all fairness, the atmosphere itself probably does a heck of a lot of that kind of thing so it's not at all unrealistic. Just annoying in a flight sim.

Unfortunately it seems that DLSS causes some motion blurring and weird artifacts on cockpit displays and mirrors. So, it's implementation in DCS is not perfected yet. Performance in the Multithreaded version of the sim is so good now that by the time the Vulkan API version comes around, I doubt it'll be necessary for many people to use DLSS at all unless you're on an uber-tight budget and need to use your old ass RTX 2060.

AD

On 1/16/2024 at 9:45 PM, twistking said:

If you have some sort of god complex, it must be truly difficult to endure the opinion of heretics.

TECHNOLOGY IS A SIN!!! ELECTRICITY IS A SIN!!!

No wait, that was a different game!

AD

[Gunfreak]

Don't use it. Except for using performance DLSS I see no useful increase in FPS and on performance it's all mosh. 

But I can fly through a b17 bomber formation with 96 bombers and 44 Germans buzzing around and its playable in vr at ludicrous resolution in my pimax. So I'm happy as it is.

[Comstedt86]

On 1/17/2024 at 7:24 AM, twistking said:

I don't have an opinion on wether that's true with the current state of the technology, but in theory (!) a technique like DLSS could archive higher quality than raw rendering even if the raw image is rendered with MSAA. A technique that uses subpixel shifting with motion vectors to combine shifted images in a temporally stable way (what DLSS does, or tries to do) can archive a significantly higher effective resolution. People tend to not understand this: Techniques like DLSS accumulate image information over subsequent frames allowing it in theory to reconstruct an image with higher detail than "ground truth". The ideal case would be a static scene with a high frame rate. In the typical case (motion scene, lower framerate) the quality of the reconstruction is highly affected by the success of the temporal matching. In practice this is what leads to the perceived blurriness or -in the worst case- ghosting.

A good practical example of this is using FSR2 or DLSS on a 1440P monitor with 4K downscaled resolution set. Running FSR or DLSS on Quality means it will render at 1440P if I remember correctly but will generally look better than native 1440P rendering but at some performance cost, sometimes even lower Quality settings can achieve good results with increased texture sharpness and anti-aliasing, but depends on the game & implementation.

[draconus]

1 hour ago, Comstedt86 said:

A good practical example of this is using FSR2 or DLSS on a 1440P monitor with 4K downscaled resolution set.

No, both techs are upscalers so they render at lower resolutions than your target resolution.

[M1Combat]

AMD's FSR is much better.  Does all the things that DLSS does visually...  but without the ghosting.  Way better...  Even on my NVidia 3080 ;)...

[Hiob]

When you want higher quality for a given resolution, you need AA without upscaling. DLAA or MSAA (all others are basically crap in comparison).

DLAA uses the same techniques and algorithms as DLSS, only in the other direction. Or in other words, without downscaling the render resolution beforehand.

[Nedum]

On 1/17/2024 at 3:45 AM, twistking said:

If you have some sort of god complex, it must be truly difficult to endure the opinion of heretics.

Yes, it is, that's for sure. Feel pleased to hear my truths.

On 1/17/2024 at 6:49 AM, Art-J said:

Then stop acting like "believer" yourself. Nobody in this thread said DLSS gives better quality than raw rendering. We all know it comes at a price of some quality. But acceptable or "noticeable" quality reduction from using it is a personal preference and different for everyone, just like acceptable minimum framerate, or acceptable level of stuttering. You seem to be blind to that fact, though.

 

Oh, so no one is of this opinion? What about the following quote (see below)?
 

On 1/17/2024 at 7:24 AM, twistking said:

I don't have an opinion on wether that's true with the current state of the technology, but in theory (!) a technique like DLSS could archive higher quality than raw rendering even if the raw image is rendered with MSAA. A technique that uses subpixel shifting with motion vectors to combine shifted images in a temporally stable way (what DLSS does, or tries to do) can archive a significantly higher effective resolution. People tend to not understand this: Techniques like DLSS accumulate image information over subsequent frames allowing it in theory to reconstruct an image with higher detail than "ground truth". The ideal case would be a static scene with a high frame rate. In the typical case (motion scene, lower framerate) the quality of the reconstruction is highly affected by the success of the temporal matching. In practice this is what leads to the perceived blurriness or -in the worst case- ghosting.

Interestingly there have been many related techniques used to enhance video capture in the real world. For example subpixel shifting for ultra high resolution still photography.

Guessing should in theory more precise than knowing? In what Universe?

DLSS subtracts pixels from an image to reduce its size. Image information is lost in the process. This information then has to be guessed when restoring the original resolution, and this has to be done every time the image is resized. The faster the image shifts its content (pixels), the more complicated the guessing process becomes, even more with overlapping fast moving content. Pixel errors and ghosting are the ugly "children" of this "marriage", and they will always be there, not always so ugly, but they are the product of omitting important "genetic material".
And as long as we can't look into the future to know what tomorrow will have been, nothing will change.
This is physics, not make a wish.

[Wali763]

Did not read through the thread, only parts. Sry.

I dont like DLSS, makes gauges rather hard to read. Also ghosting. DLAA removes all the shimmering but makes the image blurry.

Still running lower settings with 2xMSAA to achieve 90fps. No DLxx for me. For me the best compromise.

 

[Hiob]

10 minutes ago, Wali763 said:

Did not read through the thread, only parts. Sry.

I dont like DLSS, makes gauges rather hard to read. Also ghosting. DLAA removes all the shimmering but makes the image blurry.

Still running lower settings with 2xMSAA to achieve 90fps. No DLxx for me. For me the best compromise.

 

Well, when your hardware signature is correct, you can’t even use DLAA (or DLSS for that matter).

[DCoffey]

2 hours ago, Nedum said:

 

Guessing should in theory more precise than knowing? In what Universe?

 

Says all I need to know about you. Have a good day "Karen"

[MadKreator]

DLSS is prob great for lower end systems that struggle a bit. I replaced the DLSS DLL 3.5 with 3.5.1 earlier today ( that I saw in some video) and it seemed to improve ghosting a bit but nothing truly spectacular. Performance improvement fps-wise, sure, but the graphical degradation is very noticeable to me on a 48” 4K oled screen using DLSS. Msaa x4 and capped at 60 fps( personal preference) is just right for me. My opinion still is DLSS was a good addition. Not everyone needs it or benefits from it so no harm using it or not.  

Edited February 22, 2024 by MadKreator

[Comstedt86]

7 hours ago, draconus said:

No, both techs are upscalers so they render at lower resolutions than your target resolution.

I might be terrible at explaining so I'll try again.

1.1440P monitor.

2.Resolution set at 3840x2160 downscaled/supersampling.

3. Run FSR2 or DLSS at Quality (3840x2160 = 1440P internal resolution)

Giving DLSS or FSR2 more pixels to initially work with initially yields way better results.

7 hours ago, M1Combat said:

AMD's FSR is much better.  Does all the things that DLSS does visually...  but without the ghosting.  Way better...  Even on my NVidia 3080 ;)...

Makes no sense, atleast in DCS FSR also suffers from alot of ghosting.