Fidelity Super Resolution in 2021

[Rroff]

i call bs on this, like saying you can add quality back to a compressed audio file. nonsense, so how does it magically do that then? please explain in detail because id love to here it.

While the actual implementation is a bit more rudimentary currently DLSS uses a model trained against the "style" of a game - it then analyses the low resolution output for patterns which it can use the closest reference matches for to fill in the blanks using something that mimics the style of the missing data. With enough training the model produces something close enough most people won't notice.

For audio that is even harder to do because most people would tell much easier that there is nonsense noise filling in the blanks, with images it is a bit easier to fool the eyes.

 

[zx128k]

While the actual implementation is a bit more rudimentary currently DLSS uses a model trained against the "style" of a game - it then analyses the low resolution output for patterns which it can use the closest reference matches for to fill in the blanks using something that mimics the style of the missing data. With enough training the model produces something close enough most people won't notice.

For audio that is even harder to do because most people would tell much easier that there is nonsense noise filling in the blanks, with images it is a bit easier to fool the eyes.

 

[D.P.]

i call bs on this, like saying you can add quality back to a compressed audio file. nonsense, so how does it magically do that then? please explain in detail because id love to here it.

because it accumulates information across multiple frames

 

[Grim5]

audio is probably even easier for AI because of repetition, once it knows what a drum sounds like and that it's in tune with the beat it should be rather simple to add in AI assisted sound.

but it's also not the best example - heck i've tried and I can't hear any difference between MP3's and Tidal's uncompressed Hi-Res audio.

And other animals have much better hearing than humans do, while we have superior eye sight

 

[Game]

good lord is this argument about my man bits are bigger than yours still going on

 

[ICDP]

good lord is this argument about my man bits are bigger than yours still going on

Yes, every now and then they come in here to bore us and go into some pointless detail why DLSS is so much superior to FSR. The last time I posted it was about an FSR mod for all VR games and I was asked to "post screenshots to prove it works"... NO go ******* try it yourself, it's free.

 

[Game]

Yes, every now and then they come in here to bore us and go into some pointless detail why DLSS is so much superior to FSR. The last time I posted it was about an FSR mod for all VR games and I was asked to "post screenshots to prove it works"... NO go ******* try it yourself, it's free.

they are only trying to justify the 3090's

the FSR games I tried it worked and while running you couldnt tell the difference

but.. anyone can screenshot but its not what you see when playing

 

[zx128k]

Yes, every now and then they come in here to bore us and go into some pointless detail why DLSS is so much superior to FSR. The last time I posted it was about an FSR mod for all VR games and I was asked to "post screenshots to prove it works"... NO go ******* try it yourself, it's free.

Well do you blame them, after attacking DLSS with a magnifing glass for not being native quality but accepting FSR which is far worse in quality than DLSS.

 

[ICDP]

Well do you blame them, after attacking DLSS with a magnifing glass for not being native quality but accepting FSR which is far worse in quality than DLSS.

This is revisionist ******** and has already been debunked, FSR is demonstrably not "far worse in quality to DLSS". In fact most credible reviews have said FSR 1,0 is very close to DLSS 2.0 in quality and a very good first effort. DLSS 1.0 didn't need a magnifying glass to show it was crap and was rightly panned by everyone on release. FSR is far superior to what DLSS 1.0 was capable of on release.

DLSS 2.0 and later are excellent and nobody is saying FSR is technically better, just close enough in most cases. Add to this that FSR works on far more GPUs and is easier to implement, so potentially a better solution all things considered.

I don't need to compare DLSS and FSR in the same game to see FSR is objectively giving close to 4K native. Or in some cases subjectively improving over native (DOTA 2 and Terminator Resistance). The point is that FSR is good enough to pass the pixel peep test and will benefit a much wider userbase.

 

[Rroff]

audio is probably even easier for AI because of repetition, once it knows what a drum sounds like and that it's in tune with the beat it should be rather simple to add in AI assisted sound.

but it's also not the best example - heck i've tried and I can't hear any difference between MP3's and Tidal's uncompressed Hi-Res audio.

And other animals have much better hearing than humans do, while we have superior eye sight

Possibly with more advanced systems but current deep learning models don't really cut it for broad audio use vs trained to mimic detail in a specific game.

This is revisionist ******** and has already been debunked, FSR is demonstrably not "far worse in quality to DLSS". In fact most credible reviews have said FSR 1,0 is very close to DLSS 2.0 in quality and a very good first effort. DLSS 1.0 didn't need a magnifying glass to show it was crap and was rightly panned by everyone on release. FSR is far superior to what DLSS 1.0 was capable of on release.

DLSS 2.0 and later are excellent and nobody is saying FSR is technically better, just close enough in most cases. Add to this that FSR works on far more GPUs and is easier to implement, so potentially a better solution all things considered.

I don't need to compare DLSS and FSR in the same game to see FSR is objectively giving close to 4K native. Or in some cases subjectively improving over native (DOTA 2 and Terminator Resistance). The point is that FSR is good enough to pass the pixel peep test and will benefit a much wider userbase.

Nah even with 2.x DLSS certain people were applying standards to it they haven't been with FSR. It is pretty noticeable to me with both never mind pixel peeping I can see the loss of detail in certain areas a mile off with FSR and likewise see all the deficiencies of DLSS especially around motion of long thin objects.

 

[ICDP]

I try to look at each tech from an objective position. I would prefer that neither FSR or DLSS even existed and we had GPUs that would not require a crutch to get decent FPS with RT enabled. Having said that I accept that if I want RT effects I have to use DLSS or FSR to get it.

I can see the advantages and disadvantages of both DLSS and FSR and from testing both, they deliver acceptable Image Quality for the performance uplift. So when people say FSR has "far worse quality" I call them out on it because I know from my own objective testing that this is simply not the case.

I have presented 1080p comparisons in both DOTA 2 and Terminator Resistance to show that even at 1080p FSR can be implemented without the claimed "massive image blurring". The problem is when people parrot other people's opinions or reviews as facts without verifying for themselves. For example I have seen claims that DLSS adds lost detail in CP2077 and when you look at some "native vs DLSS" images you can indeed see that DLSS seems to add missing detail. But that ignores the fact that CP2077 forces TAA which adds blur to the native image. So it's not a level playing field in a lot of these tests.

 

[Kevin Tanton]

FSR has been added to the custom shaders patch for Assetto Corsa. Looks good, some people are so clever.

 

[ICDP]

FSR has been added to the custom shaders patch for Assetto Corsa. Looks good, some people are so clever.

The problem with these FSR mods is that they also affect the menu UI items. So any menu text/icons are also rendered at a lower resolution and upscaled. FSR is specifically meant to be implented earlier in the rendering pipeline before any IU assets are added. So the UI is always rendered at native and not blured and or sharpened.

Having said that I have tested the FSR Open VR mod in IL-2, DCS, Fallout 4VR and Skyrim VR and saw a notible performance increase for very little IQ loss.

 

[zx128k]

This is revisionist ******** and has already been debunked, FSR is demonstrably not "far worse in quality to DLSS". In fact most credible reviews have said FSR 1,0 is very close to DLSS 2.0 in quality and a very good first effort. DLSS 1.0 didn't need a magnifying glass to show it was crap and was rightly panned by everyone on release. FSR is far superior to what DLSS 1.0 was capable of on release.

DLSS 2.0 and later are excellent and nobody is saying FSR is technically better, just close enough in most cases. Add to this that FSR works on far more GPUs and is easier to implement, so potentially a better solution all things considered.

I don't need to compare DLSS and FSR in the same game to see FSR is objectively giving close to 4K native. Or in some cases subjectively improving over native (DOTA 2 and Terminator Resistance). The point is that FSR is good enough to pass the pixel peep test and will benefit a much wider userbase.

FSR is just upscaling. It uses the same sparial upscale as DLSS 1, then edge reconstruction and sharpening. Its never going to be good quality. Only AI upscaling is good quality, like Gigapixel AI. FSR is 100% the same method as DLSS 1 and AMD have a patent to add an AI network to a FSR upscaling method. The main difference that makes DLSS 1 better is the AI network. This adds back some of the detail FSR cannot recreate.

FSR cannot add anything to the internal render image. Thus DLSS is better as it adds detail back. Hair detail for example is blurred out of existance by FSR but the detail is restored by DLSS. FSR can never match the native resolution by design. Basically the internal lower resolution remains in the upscaled image in every texture. Details present at the higher native resolution are basically gone in the FSR image. This means the image is always blurred reguardless of the quality setting. There is nothing that can be done about it, its how FSR works. You can argue untill you are blue in the face but FSR always reduces image quality and never is as detailed as native. Its how the tech works, there is no way to win a debate stating otherwise.

All versions of DLSS are better by design. DLSS 2 far better because they no long use the same method as FSR and DLSS to upscale. DLSS 2 uses tempoeral upscaling like Unreal Engine 5 and motion vectors. This lead to the increase in quality in DLSS 1.9 and the versions above.

FSR is a simple, fast upscaling method. DLSS is a complex image reconstruction method.

 

[ICDP]

Snip

More opinion masquerading as facts and nobody gives a crap who they work, just that they do. I'm going to go ahead and believe my own testing results and the conslusions from reputable reviews before the random ramblings by some "self-appointed" expert on upscaling.

I even posted images showing that FSR can and does give increased detail when implemented correctly, even at 1080p. But you go ahead spouting your technical babble as if it means something.

Oh look, FSR doing what you said is impossible and adding more detail in some cases.

Spoiler: FSR 75% vs Native

 

[Zarax]

1st one is FSR?

 

[ICDP]

1st one is FSR?

2nd one is FSR at 75% at 1080p, if you right click and open in a new tab the image names show. It will of course be subjective but to me it adds some extra details around the saddlebags, rocks and stonework on the ground as well as giving better anti-aliasing. So I prefer the FSR image over the native image but can see where it would be subjective. My point is that contrary to what some self-appointed expert telling us what FSR cannot do, it's doing the impossible and adding detail.

Bear in mind this is a native 1080p image vs a 810p image with FSR added. I don't care if it's "only" upscaled and sharpened and not pure wonderful DLSS deep AI learning witchcraft. I only care that it works and gives better performance for marginal IQ loss, or arguably equal or better IQ in some cases.

Like I said I follow the evidence and my own testing, not some so called expert telling me that what I am seeing is actually impossible.

 

[Game]

Well do you blame them, after attacking DLSS with a magnifing glass for not being native quality but accepting FSR which is far worse in quality than DLSS.

might as well close this thread its very little to do with the subject

 

[Zarax]

2nd one is FSR at 75% at 1080p, if you right click and open in a new tab the image names show. It will of course be subjective but to me it adds some extra details around the saddlebags, rocks and stonework on the ground as well as giving better anti-aliasing. So I prefer the FSR image over the native image but can see where it would be subjective. My point is that contrary to what some self-appointed expert telling us what FSR cannot do, it's doing the impossible and adding detail.

Bear in mind this is a native 1080p image vs a 810p image with FSR added. I don't care if it's "only" upscaled and sharpened and not pure wonderful DLSS deep AI learning witchcraft. I only care that it works and gives better performance for marginal IQ loss, or arguably equal or better IQ in some cases.

Like I said I follow the evidence and my own testing, not some so called expert telling me that what I am seeing is actually impossible.

Interesting. I wanted to actually guess, looks like the kind of distortion created by the reconstruction can actually be taken as extra detail in some occasions.

 

[PrincessFrosty]

All versions of DLSS are better by design. DLSS 2 far better because they no long use the same method as FSR and DLSS to upscale. DLSS 2 uses tempoeral upscaling like Unreal Engine 5 and motion vectors. This lead to the increase in quality in DLSS 1.9 and the versions above.

FSR is a simple, fast upscaling method. DLSS is a complex image reconstruction method.

It's worth highlighting this point because they are fundamentally different systems although they appear similar at first. It's interesting to me because the use case for DLSS almost all of the time right now is fundamentally to enable the use of ray tracing, since that just cannot run at high resolution. And so a real life use case for me was RT enabled in Cyberpunk with DLSS to get up to 4k, and if we could compare the quality of FSR to DLSS in that kind of use case then yes absolutely, the quality loss for FSR from 1080p to 4k is going to be absolutely awful.

Of course quality modes look much nicer because you're starting with a much higher internal resolution, which is fine if you want to make the quality/speed tradeoff just for more fps...but I don't think that's really a use case for DLSS. I'd be interested to know if anyone is using DLSS in games for better frame rates without the use of RT? My guess is that people generally still have a preference for native, but I guess we'd need data on that. I'm more of stickler for image quality however, I don't even really like FXAA or even really TAA, I like those crisp clean textures and to lower that for a few fps is a hard sell. If I'm getting something truly new and unique like RT effects then it's different story. Especially now DLSS is looking so good.