DLSS 4.0 Super Resolution Stress Test: Does The Transformer Model Fix The Biggest Issues?

Bojji

Member


Conclusion:

-It fixes some things
-It doesn't fix others
-It also can introduce new issues

Biggest difference is probably clarity in motion:

z01wnSz.jpeg


In my experience DLSS4 is generally better than DLSS3 but there are instances where DLSS3 produces better image.
 
Last edited:
dlss 4 is true magic
I was already fine with dlss 3 performance at 1440p (720p internal). it looked better than 1080p already
now 1440p dlss 4 performance looks better than 1440p itself which is an incredible boost to performance for my 3070
 
Stress testing the 'ever so popular' 1440p resolution. In 2025. In an era when cards are more expensive than before.

Im sorry but the video he shouldve made was that no one in 2025 should be playing on 1440p with internal resolutions as low as 720p.

Ive been gaming at 4k for the last 9 years now. its shocking to see standards become so low just to hold 60 fps. And people like Alex are not helping promoting videos like this.
 
Stress testing the 'ever so popular' 1440p resolution. In 2025. In an era when cards are more expensive than before.

Im sorry but the video he shouldve made was that no one in 2025 should be playing on 1440p with internal resolutions as low as 720p.

Ive been gaming at 4k for the last 9 years now. its shocking to see standards become so low just to hold 60 fps. And people like Alex are not helping promoting videos like this.

Its a better stress test of the technology than 4k, simple as that. The lower the internal resolution the more you test how the good the upscaler is.
 
Its a better stress test of the technology than 4k, simple as that. The lower the internal resolution the more you test how the good the upscaler is.
Then the parameters need to be changed. The cost of DLSS has increased from 1 to 2 to now DLSS4. While 4k tv and monitor adoption has only increased since 2018. They should instead of doing more with less. Not the other way around. whats the point of adding more and more tensor cores if the cost keeps on increasing?
 
Last edited:
DLSS4 is indeed quite good, but since I game at 4K, DLSS3 (heck even FSR3) was already amazing and I cant say that the improvement is drastic, except that I can now run in performance mode and not notice anything being truly lost. DLSS4 did lose some performance, but the fact that you can drop it to balanced or performance to make up the difference is truly incredible.

However, as incredible as DLSS4 is, I cannot get over how much better FSR now is. DLSS4 still beats FSR4, but it's not a worth paying a premium for anymore if your favorite games support both (which most dont). It went from having mediocre to bad (unless you were gaming at 4K) to outright amazing. Easily the biggest jump in quality we have seen since DLSS1 (which was AWFUL) to DLSS2.
 
Last edited:
Then the parameters need to be changed. The cost of DLSS has increased from 1 to 2 to now DLSS4. While 4k tv and monitor adoption has only increased since 2018. They should instead of doing more with less. Not the other way around. whats the point of adding more and more tensor cores if the cost keeps on increasing?

Tom Cruise What GIF


DLSS 4 has relatively low cost of TOPs, goes from a 5090 with a slight advantage in performance penalty like Ada, but a few % increase on Ampere and Turing, all the way down to a 2050 laptop. That means that even throwing more tensor cores or TOPs at it is not making a miracle that it needs a certain number of milliseconds to inference the image. Even if you threw all the 5090 TOPs at the problem, who's to say that in the queue of image handling you're simply limited?

Those extra TOPs are for ray reconstruction, neural radiance cache path tracing, future neural texture decompression, neural this neural that, etc.

DLSS 4 performance can sometimes look better than DLAA CNN. Really not sure why you're aggro on this. Dude, you're on Ampere, you get a massive boost of performance with DLSS 4 by having the lower quality tiers actually look good and not trash like it was before under quality.
 
Last edited:
The lowest I do is 1080p performance to 4k upscaled dlss.

I wonder how many people realise the resolution they're actually playing at on lower resolution screens like 1440p and 1080p.

75zrcIL.jpeg
 
Extremely disappointed at it introducing some issues. You'd think with how much heavier it is, it would utterly crush DLSS3 and get rid of most problems, but nah.
 
Last edited:
Those extra TOPs are for ray reconstruction, neural radiance cache path tracing, future neural texture decompression, neural this neural that, etc.
So future shit we cant use right now. yay? I am saying the focus should be on reducing the cost of DLSS so more people can use 4k in 2025 instead of being forced to slum around in 1440p with low 720p internal resolutions just to hit a solid 60 fps.
DLSS 4 performance can sometimes look better than DLAA CNN. Really not sure why you're aggro on this. Dude, you're on Ampere, you get a massive boost of performance with DLSS 4 by having the lower quality tiers actually look good and not trash like it was before under quality.

DLSS4 performance does not even look as good as DLSS quality. i dont know where you are getting this from. this video alone shows there are issues still with the DLSS and should not be used for every game.

for me personally, i can still tell that dlss4 4k performance is softer than dlss3 quality. Plus the performance hit is too high as you can see below with performance being roughly equivalent to dlss3 balanced. Not to mention all the issues Alex covered above.

SytZKn7.jpeg


the funny thing is i remember dlss being sold as a performance improvement tool which would give you native quality and far better performance thanks to the GPU being freed up from having to render native pixels. 7 years later, the performance has only gotten worse despite all the technological advancements, and we are essentially celebrating 1440p being the most popular resolution.
 
The lowest I do is 1080p performance to 4k upscaled dlss.

I wonder how many people realise the resolution they're actually playing at on lower resolution screens like 1440p and 1080p.

75zrcIL.jpeg
native taa / dlss 3 looks so blurry at 1080p and 1440p that it stops mattering, really

witcher 3 at 1440p dlss 3 quality (960p) vs. dlss 4 performance (720p)

cyberpunk at native 1080p dlss 3 vs. 1080p dlss 4 quality (720p)

marvel rivals at native 1080p taa vs. 1080p dlss 4 performance (540p)


it has its issues but for people like me who likes to have motion clarity but can't play at 4K, dlss 4 is a game changer
 
Last edited:
So future shit we cant use right now. yay? I am saying the focus should be on reducing the cost of DLSS so more people can use 4k in 2025 instead of being forced to slum around in 1440p with low 720p internal resolutions just to hit a solid 60 fps.

Slim, there's clearly a hard limit to the queue and inference time no matter what you throw at it. How much you think you can optimize from ~1-2 ms ? Nobody has found a solution to near zero latency. DLSS 4 is like a what, ~0.5ms additional? If even that. You have a perception of what that is in a game pipeline?

Not sure either what you're babbling about, peoples with 4k monitors will surely use 4k resolution. Nvidia gave DLSS 4 performance for those peoples, because yes, it does look better, as we'll see..

DLSS4 performance does not even look as good as DLSS quality. i dont know where you are getting this from. this video alone shows there are issues still with the DLSS and should not be used for every game.

Hard disagree and has been shown to be a LOT of times. Texture details at performance beat DLSS CNN quality and even DLAA sometimes. In motion its even better.



Don't tell me you go out of your way now to return to CNN model?

for me personally, i can still tell that dlss4 4k performance is softer than dlss3 quality. Plus the performance hit is too high as you can see below with performance being roughly equivalent to dlss3 balanced. Not to mention all the issues Alex covered above.

SytZKn7.jpeg

Why not include FSR 4? I know exactly where you picked it

Cxdngdw.jpeg


Seems like there's no such thing as a free lunch with upscalers and performances hit.

Alex also showcases a lot of issues better resolved and if you look at his past video, the quality is much better. Even AMD unboxed says its mind blowing. Not to mention in motion. If you can't see the difference in motion then I don't know what to tell you. Ain't no way I'm saving a tiny % for returning back to CNN. Quoting Tim "If DLSS 4 represents 4k with ultra textures, DLSS 3 the old version is more like 1440p with high textures in comparison.

the funny thing is i remember dlss being sold as a performance improvement tool which would give you native quality and far better performance thanks to the GPU being freed up from having to render native pixels. 7 years later, the performance has only gotten worse despite all the technological advancements, and we are essentially celebrating 1440p being the most popular resolution.

Are you really complaining that 1440p is getting more popular and peoples aren't picking up 4k? If anything, 1440p is the real sweet spot imo for PC monitors. Its also tougher on upscalers. 4K will have less of these problems showcased in the DF video. Hurray for you? You complain about performances but you picked 4k over 1440p, do you not see the irony?

Go back to native or CNN for all I care. Blows my mind anyone with Nvidia card would subject himself to not use it.

rr1btsygj0fe1.jpeg
 
Last edited:
Best game for me to test it is Dragon's Dogma 2. The water looks much better with DLSS4 and there's very little artifacting, generally.

It's what the game should have been like at launch.
 
Stress testing the 'ever so popular' 1440p resolution. In 2025. In an era when cards are more expensive than before.

Im sorry but the video he shouldve made was that no one in 2025 should be playing on 1440p with internal resolutions as low as 720p.

Ive been gaming at 4k for the last 9 years now. its shocking to see standards become so low just to hold 60 fps. And people like Alex are not helping promoting videos like this.
1440p offers the perfect blend of high res and high fps for 32" and below. On my card 1440p runs 70% faster than 4K. Much better motion clarity and input latency for a generally negligible drop image quality.
 
Yeah been forcing dlss 4 on all the games that can support it. It is really something special. I run native 1440p on some old gen games and its not even close to dlss 4 quality at the same res.
 
Last edited:
Yeah been forcing dlss 4 on all the games that can support it. It is really something special. I run native 1440p on some old gen games and its not even close to dlss 4 quality at the same res.

Me too but it's such a pain in the ass. There has to be a better way than manually changing it in the Nvidia app for every game.
 
I would always avoid using 1440p balanced mode (let alone performance mode) unless absolutely necessary.

at 1440p, quality mode is great, but anything below that becomes problematic, unless the game has some really bad TAA.
Balanced mode is only really good at 4k or other markedly higher resolutions where the base resolution doesn't go below 900p.

that's IMO a general rule for DLSS.
if you want actually good image quality, make sure the internal resolution isn't below 900p... preferably close to 1080p

1440p Quality Mode = 960p 👍
1440p Balanced Mode = 835p 👎🏼
 
Last edited:
I would always avoid using 1440p balanced mode (let alone performance mode) unless absolutely necessary.

at 1440p, quality mode is great, but anything below that becomes problematic, unless the game has some really bad TAA.
Balanced mode is only really good at 4k or other markedly higher resolutions where the base resolution doesn't go below 900p
that's the point, dlss 4 performance looks better than dlss 3 quality

if you were fine with 1440p dlss 3 quality, then you should be fine with 1440p dlss 4 performance
 
that's the point, dlss 4 performance looks better than dlss 3 quality

if you were fine with 1440p dlss 3 quality, then you should be fine with 1440p dlss 4 performance

I haven't tested it yet much tbh, but in motion I would expect noticeably more artifacting still. but I might actually try forcing it in games that I know well to see.
 
I haven't tested it yet much tbh, but in motion I would expect noticeably more artifacting still. but I might actually try forcing it in games that I know well to see.
this is in motion, hence they're not properly aligned (because it is really not easy)
 
this is in motion, hence they're not properly aligned (because it is really not easy)

what kind of motion?
a stress test would be strafing to the side with parallaxing objects. that way you can judge possible disocclusion and ghosting issues the best I think.

like when I compared Indiana Jones with native TAA vs DLSS Quality at 1440p in another thead, to show DLSS has superior image quality.

there I strafed to the left with trees in the background, a wall in front of me, and a window with thin metal bars in the center. you could really nicely compare the ghosting and image sharpness in that image I think.
 
Last edited:
what kind of motion?
a stress test would be strafing to the side with parallaxing objects. that way you can judge possible disocclusion and ghosting issues the best I think
well can't do that sadly, this one took a long time already
it is sprinting back and forth

anyways any ghosting issues I'm seeing with dlss 3 or dlss 4 are not resolved with the internal resolution. see ac shadows, dlss 4 has particle ghosting no matter quality, balanced or performance mode
 
Last edited:
I'll do one later maybe if I don't forget.

I still have Avowed installed. what did you use to force it? does the Nvidia app detect it correctly?
yes nvidia app detects it correctly on my end (game pass version)

don't get me wrong dlss quality still looks better than dlss performance with dlss 4. but I'd prefer dlss 4 performance over dlss 3 quality

and that's where the problem is actually, I never was actually happy with 1440p dlss 3 quality. it always looked blurry to me. that is why I like dlss 4 so much
 
Last edited:
yes nvidia app detects it correctly on my end (game pass version)

don't get me wrong dlss quality still looks better than dlss performance with dlss 4. but I'd prefer dlss 4 performance over dlss 3 quality

and that's where the problem is actually, I never was actually happy with 1440p dlss 3 quality. it always looked blurry to me. that is why I like dlss 4 so much

it depends highly on the game. I have demonstrated in another thread before in Indiana Jones that 1440p DLSS 3 Quality looks objectively better than its native TAA setting.
it had less ghosting, less artifacts and a sharper look generally.

here are the comparison images (I am strafing left, so look for artifacts and sharpness between the metal bars, the cobwebs, the trees etc.)

dyzfcveo.png


ah9qyjjt.png


it should be obvious which one is which due to the FPS counter top right lol.

the metal bars show excessive ghosting with TAA
 
Last edited:
that's the point, dlss 4 performance looks better than dlss 3 quality

if you were fine with 1440p dlss 3 quality, then you should be fine with 1440p dlss 4 performance

I'M BACK xD
ok, so I made 4 images of Avowed. DLSS 3 and DLSS 4, each in Performance and in Quality mode.
I have hidden the performance on the top right and named the files 1, 2, 3, 4... I wrote down which one is which in a text file so I don't forget.

now I wonder how people like these.
they are all targeting 1440p, all are png files.
I am strafing to the left, and deliberately have a wooden structure in front of me parallaxing with the background, the gun will of course also exhibit artifacting to the right of it as the background parallaxes with it.

1
k273xsni.png


2
numyxp3t.png


3
mr8y5fbt.png


4
otaqlfob.png



doing these comparisons, I am once again remembered just how FUCKING ABSOLUTELY UGLY AND DOOOOOOGSHIT UE5's Lumen is... my god... it's the worst RT implementation out there...
sadly you can't get that across in screenshots, even in motion, as the boiling and constant noise of it doesn't really come across when it's not in motion... it's so ass... LUMEN IS CANCER.

reveal in:
Apr 16, 2025 at 12:00 AM
 
Last edited:
my guess

1st - dlss 4 quality
2nd - dlss 3 performance
3rd - dlss 3 quality
4th - dlss 4 performance

yeah nanite and lumen, this game has it all but it looks so much worse than ac shadows and indiana jones for me. i don't know why. I didn't notice the boiling and noise because I'm using supersampling. I also heard that you need to set some settings to epic if you don't want noise if you're going to use upscalers. for example reflections are really broken with anything below epic when you use upscaling
 
Last edited:
I'M BACK xD
ok, so I made 4 images of Avowed. DLSS 3 and DLSS 4, each in Performance and in Quality mode.
I have hidden the performance on the top right and named the files 1, 2, 3, 4... I wrote down which one is which in a text file so I don't forget.

now I wonder how people like these.
they are all targeting 1440p, all are png files.
I am strafing to the left, and deliberately have a wooden structure in front of me parallaxing with the background, the gun will of course also exhibit artifacting to the right of it as the background parallaxes with it.

1
k273xsni.png


2
numyxp3t.png


3
mr8y5fbt.png


4
otaqlfob.png



doing these comparisons, I am once again remembered just how FUCKING ABSOLUTELY UGLY AND DOOOOOOGSHIT UE5's Lumen is... my god... it's the worst RT implementation out there...
sadly you can't get that across in screenshots, even in motion, as the boiling and constant noise of it doesn't really come across when it's not in motion... it's so ass... LUMEN IS CANCER.
#1 and #4 look best (#1 looks ever so slightly better)
#2 looks worst
 
I'M BACK xD
ok, so I made 4 images of Avowed. DLSS 3 and DLSS 4, each in Performance and in Quality mode.
I have hidden the performance on the top right and named the files 1, 2, 3, 4... I wrote down which one is which in a text file so I don't forget.

now I wonder how people like these.
they are all targeting 1440p, all are png files.
I am strafing to the left, and deliberately have a wooden structure in front of me parallaxing with the background, the gun will of course also exhibit artifacting to the right of it as the background parallaxes with it.

1
k273xsni.png


2
numyxp3t.png


3
mr8y5fbt.png


4
otaqlfob.png



doing these comparisons, I am once again remembered just how FUCKING ABSOLUTELY UGLY AND DOOOOOOGSHIT UE5's Lumen is... my god... it's the worst RT implementation out there...
sadly you can't get that across in screenshots, even in motion, as the boiling and constant noise of it doesn't really come across when it's not in motion... it's so ass... LUMEN IS CANCER.

reveal in:
Apr 16, 2025 at 12:00 AM

so the countdown is up, there is the reveal of which is which:

1 = DLSS 4 Quality
2 = DLSS 3 Performance
3 = DLSS 3 Quality
4 = DLSS 4 Performance

yamaci17 yamaci17 was dead on


here are the uncensored images in the same order but with the performance data uncovered:

DLSS 4 Quality
h8qlakvu.png


DLSS 3 Performance
r5z5j3pw.png


DLSS 3 Quality
5ztue3ag.png


DLSS 4 Performance
2j3umpce.png



as you can see, DLSS4 has a 3 to 4 FPS penalty over DLSS3 in each mode here.
one thing the DLSS4 Performance can not get to the level of DLSS3 Quality is the raytracing.
the low internal resolution means a lower ray count for the already ugly as fuck Lumen RT. this means reflections look blocky as hell, and the lighting is boiling like a motherfucker.

so in UE5, when actually playing, DLSS4 Performance looks less appealing than DLSS 3 Quality. in games with no raytracing it would look superior throughout I assume.
 
Last edited:
as you can see, DLSS4 has a 3 to 4 FPS penalty over DLSS3 in each mode here.
one thing the DLSS4 Performance can not get to the level of DLSS3 Quality is the raytracing.
the low internal resolution means a lower ray count for the already ugly as fuck Lumen RT. this means reflections look blocky as hell, and the lighting is boiling like a motherfucker.

so in UE5, when actually playing, DLSS4 Performance looks less appealing than DLSS 3 Quality. in games with no raytracing it would look superior throughout I assume.
you can fix that by using epic global illumination and reflection
but then that could defeat the purpose of using dlss performance to begin with
so you have a point there
then again this was not a problem with cyberpunk and ray tracing so I don't know what to say
 
Last edited:
Image quality is great but it had weird ghosting in some area . Characters stand infront of grass > moving camera > fsr3.0 tier ghosting .
 
Interesting; the DLSS4 "limitations" remind me a lot the PSSR flaws.
Has somebody tried AC shadows to do comparisons of PSSR vs DLSS3 vs DLSS4 with same native resolution?

Last time DF did a PSSR vs DLSS3 comparison with same native resolution, PSSR was handily beating DLSS3 (like it was not even a contest). From there they stopped doing those fair comparisons AFAIK.
 
Has somebody tried AC shadows to do comparisons of PSSR vs DLSS3 vs DLSS4 with same native resolution?

Last time DF did a PSSR vs DLSS3 comparison with same native resolution, PSSR was handily beating DLSS3 (like it was not even a contest). From there they stopped doing those fair comparisons AFAIK.

Why am I not surprised?
 
Has somebody tried AC shadows to do comparisons of PSSR vs DLSS3 vs DLSS4 with same native resolution?

Last time DF did a PSSR vs DLSS3 comparison with same native resolution, PSSR was handily beating DLSS3 (like it was not even a contest). From there they stopped doing those fair comparisons AFAIK.
When was that?
 
When was that?
Jedi Survivor I think. The PSSR image had higher detail in motion. But it ignores the fact it was DLSS 3.1 model being used to compare (not 3.7), DLSS has the sharpening pass completely disabled (with no way to enable it in-game), and that the PSSR image still has worse temporal stability. The "film-grain" like effect is still present. PSSR handily beating DLSS is only true if you don't bother adding the latest DLSS 3 version, ignore sharpening entirely, and rely entirely on .gifs instead of actually comparing the two yourself.
 
Last edited:
Why am I not surprised?

When was that?
Here. Just ignore the same biased bullshit : "DLSS is better here and here" while it's barely noticeable in their footage. In motion DLSS looks constantly like vaseline compared to PSSR. A big win overal for PSSR even if they don't want to admit it. They have been on the same narrative from now on, focusing on DLSS few advantages, mostly in stills, while completely ignoring the big elephant in the room: it looks blurrier in motion (using same native resolution).



Jedi Survivor I think. The PSSR image had higher detail in motion. But it ignores the fact it was DLSS 3.1 model being used to compare (not 3.7), DLSS has the sharpening pass completely disabled (with no way to enable it in-game), and that the PSSR image still has worse temporal stability. The "film-grain" like effect is still present. PSSR handily beating DLSS is only true if you don't bother adding the latest DLSS 3 version, ignore sharpening entirely, and rely entirely on .gifs instead of actually comparing the two yourself.
But that was an old PSSR version too, same argument could be applied. We know PSSR hugely improved in the areas you mentionned. There is no film grain anymore in AC Shadows for instance. This is why I am saying somebody needs to use AC shadows on PS5 Pro vs DLSS3.7 and DLSS4
 
Last edited:
Top Bottom