Discussion TPU: DLSS tested in Final Fantasy

[BFG10K]

https://www.techpowerup.com/reviews/NVIDIA/DLSS_Super_Sampling_AA_Tested_Final_Fantasy/

As if we needed more evidence of the disaster that is Turing.

This glorified upscaler DLSS is a blurry jagged mess in comparison to TAA @ 4K. Also at the same 1440p resolution, TAA is far faster given it's basically free to enable.

Oh, and DLSS only works at exactly 4K resolution in this game. So much for "deep learning" and "AI".

Epic fail.

 

[antihelten]

This glorified upscaler DLSS is a blurry jagged mess in comparison to TAA @ 4K.

Yeah it doesn't look good:

DLSS:

4K TAA:

Also at the same 1440p resolution, TAA is far faster given it's basically free to enable.

To be fair TAA at 1440P also arguably looks worse than DLSS.

1440P TAA:

So basically DLSS is a choice of ~35% better performance at the cost of significant blurring when compared to 4K TAA, or a significantly cleaner image at the cost of ~25% performance when compared to 1440P TAA.

The interesting thing would be to see how it would compare if it was possible to normalize for either performance or image quality (by using a custom resolution along with TAA).

 

[NTMBK]

Is it possible to just open the full-res original images? I think the scaling that the website performs on the image isn't helping at all.

 

[coercitiv]

The interesting thing would be to see how it would compare if it was possible to normalize for either performance or image quality (by using a custom resolution along with TAA).

HW Unboxed did that and found that 1800p TAA matches 4k DLSS in both image quality and performance. (in the Infiltrator demo which had better TAA implementation than FF XV).

 

[Cyruskain]

https://www.techpowerup.com/reviews/NVIDIA/DLSS_Super_Sampling_AA_Tested_Final_Fantasy/

As if we needed more evidence of the disaster that is Turing.

This glorified upscaler DLSS is a blurry jagged mess in comparison to TAA @ 4K. Also at the same 1440p resolution, TAA is far faster given it's basically free to enable.

Oh, and DLSS only works at exactly 4K resolution in this game. So much for "deep learning" and "AI".

Epic fail.

Maybe my eyes are bad, or my brain is wired incorrectly, but I saw a noticable improvement in image quality when using DLSS vs. TAA in the samples provided by techpowerup except for when FPS dropped below 60. It looks better to me. /shrug.

 

[amenx]

Maybe my eyes are bad, or my brain is wired incorrectly, but I saw a noticable improvement in image quality when using DLSS vs. TAA in the samples provided by techpowerup except for when FPS dropped below 60. It looks better to me. /shrug.

Same here. It looked sharper when moving slider, nothing like the images above.

 

[Dribble]

I note the conclusion from the article wasn't epic fail. Is the 30% fps gain worth the loss in image quality? - The articles answer was yes if you don't have the fps. That was also the conclusion in the comments of people with the game and a 2080Ti and 4k.

It also gets rid of motion artifacts that TAA is giving and you can't see in a static screen shot which might be more noticeable then the sort of loss of quality you can only see in zoomed in static screen shots.

Further analysis is needed. If you need the fps then 4k TAA is out, so it would make more sense to do a 4k comparison at the same fps as DLSS gives you. i.e. go through the various methods of lowering settings so you == DLSS's fps and see which one looks best.

As an aside it would be nice if these forums could show some balance - something doesn't have to be all good or all bad you know. It can have pro's and con's. DLSS is an example of that. So much emotional investment in hating something means a complete loss of ability to give a balanced analysis of anything - something I would have hoped was easier for a bunch of geeks who should be a pretty logical people.

 

[Despoiler]

It looks like DLSS does some things better in terms of IQ and some things worse than existing methods. Clearly it's faster. Though it's been shown you can do the same upscaling quality with 1800p TAA. I think the real question is still if all of this RTX tech was worth the die area and cost vs putting in more raster horsepower. I don't see this as a compelling use given there is one game, increased cost, and ways to replicate the performance using existing methods.

 

[Tup3x]

Maybe my eyes are bad, or my brain is wired incorrectly, but I saw a noticable improvement in image quality when using DLSS vs. TAA in the samples provided by techpowerup except for when FPS dropped below 60. It looks better to me. /shrug.

Basic TAA can look really nice when there's no movement. Things fall apart when things start to move. DLSS has an advantage there.

 

[antihelten]

HW Unboxed did that and found that 1800p TAA matches 4k DLSS in both image quality and performance. (in the Infiltrator demo which had better TAA implementation than FF XV).

Nice find. Of course the Infiltrator demo is arguably the best case scenario for DLSS, being a canned demo (and thus easier to train a NN model for).

Either way though, this would indicate that DLSS brings basically nothing to the table that couldn't already be achieved with TAA.

Is it possible to just open the full-res original images? I think the scaling that the website performs on the image isn't helping at all.

Here you go:

DLSS
4K TAA
1440P TAA
1440P no AA

Maybe my eyes are bad, or my brain is wired incorrectly, but I saw a noticable improvement in image quality when using DLSS vs. TAA in the samples provided by techpowerup except for when FPS dropped below 60. It looks better to me. /shrug.

Same here. It looked sharper when moving slider, nothing like the images above.

If you guys are not using 4K monitors then you have to make sure to zoom in on the images to properly compare, or alternatively look at the full-res images linked above.

 

[amenx]

If you guys are not using 4K monitors then you have to make sure to zoom in on the images to properly compare, or alternatively look at the full-res images linked above.

I'm using a 4k display, but missed the fact that you can zoom in on the pics. Without zooming in, the majority of the DLSS examples look better, clearer, at first sight. But when you zoom in, on some of the pics its the opposite, the DLSS is more blurry, esp building textures (in particular the first and last pics). DLSS vs FXAA (pic 2) clearly in favor of DLSS. Also pic 5, DLSS (67fps) vs 1440p TAA (88fps) again clearly favors DLSS. So mixed results when zooming in. I suspect when at normal viewing distance, when you're not scrutinizing the textures up close in game, DLSS would come out ahead.

 

[coercitiv]

I suspect when at normal viewing distance, when you're not scrutinizing the textures up close in game, DLSS would come out ahead.

Based on what criteria?

 

[Cyruskain]

If you guys are not using 4K monitors then you have to make sure to zoom in on the images to properly compare, or alternatively look at the full-res images linked above.

I was viewing the aforementioned comparisons at work on a 1080p standard monitor. I came home to my 4k monitor, my Asus RoG PG27UQ, and boy are you right. When zooming in closely, DLSS does muddy the graphics on some of these, although it does slightly help on others. However, my argument is that in a video game you aren't going to be pausing the game, snapshotting, and then zooming in to see what is "really" there. From a normal viewing distance, a normal gaming distance, DLSS appears to make the image nicer (in my opinion) although is it *really* making it nicer? Debateable (yes and no, as we've seen). I think it's doing exactly what it's supposed to be doing: Making it "look" like it's higher quality when it really isn't.

 

[amenx]

Based on what criteria?

Based on the slides comparison. You have to zoom in quite a bit to spot the DLSS blur on some of them, something which does not happen in-game. antihelten's pics showing the blur are zoomed in past the games viewing point, and only on those selected pics. Other pics (2 & 5) clearly favor DLS, even when zoomed in. I could have done the same as antihelten and cherry picked those to make an argument of superior DLSS results. So mixed results as I see it (from our rather limited vantage points of examining pics).

 

[coercitiv]

From a normal viewing distance, a normal gaming distance, DLSS appears to make the image nicer (in my opinion)

Making it "look" like it's higher quality when it really isn't.

I'm starting to think people confuse overall lighting, contrast and saturation with image quality, otherwise I don't see how one can claim, at the same time, that image detail is not relevant at normal viewing distance yet one of the AA methods looks "clearer" or "nicer".

Based on the slides comparison. You have to zoom in quite a bit to spot the DLSS blur on some of them, something which does not happen in-game. antihelten's pics showing the blur are zoomed in past the games viewing point, and only on those selected pics. Other pics (2 & 5) clearly favor DLS, even when zoomed in. I could have done the same as antihelten and cherry picked those to make an argument of superior DLSS results. So mixed results as I see it (from our rather limited vantage points of examining pics).

But you weren't using the zoomed details as basis for the comparison, you were actually using a scenario that ignores such details:

I suspect when at normal viewing distance, when you're not scrutinizing the textures up close in game, DLSS would come out ahead.

So, if we're looking at the game from a normal viewing distance and, according to your own assessment, we're incapable of spotting blurred details, what objective metric makes DLSS come out ahead of TAA?

What makes 4k TAA better than 1800p TAA if we can't spot the details from a normal viewing distance and why are people insisting on running native 4k?!

 

[happy medium]

Possible slight image quality degradation that can be seen best with zoomed in, with still pictures vs 35% faster gaming? Give me DLSS everyday of the week and on Sundays.

 

[BFG10K]

As an aside it would be nice if these forums could show some balance - something doesn't have to be all good or all bad you know. It can have pro's and con's. DLSS is an example of that. So much emotional investment in hating something means a complete loss of ability to give a balanced analysis of anything - something I would have hoped was easier for a bunch of geeks who should be a pretty logical people.

DLSS was one of the two "killer" features that justified the ridiculous die size and price. Raytracing is a failure and DLSS is an even bigger failure.

Any $25 Bluray player can upscale images. And given it only works at 4K, this suggests the "learning" might have to be done on a per-resolution basis, which makes it even more useless. Imagine if PhysX only worked at 1440p for example?

DLSS is a solution looking for a problem. If you need performance, 1440p + TAA is faster. If you need quality, 4K + TAA look better. And neither need a $1200 Turding.

As an aside, we already have a gold standard AA: the TAA as implemented in the Unreal 4 engine that virtually all its games inherit automatically. The image quality is amazing, akin to an offline render. And performance is free. You literally lose only 1-2 FPS in every case I've tested, even at 5K resolutions.

 

[Qwertilot]

Its a vastly better upscaler than a 25 dollar DVD!

All this image quality argument is known pretty much futile without single blinded testing.
(Especially when some people have genuine agendas either way.).

The neural nets probably do need training separately for each resolution. Even if theoretically possible to generalise I'm pretty sure you'd lose a lot of image quality.

 

[coercitiv]

All this image quality argument is known pretty much futile without single blinded testing.
(Especially when some people have genuine agendas either way.).

It's even funnier if you consider the obvious consequence: assuming we all openly agree 4K DLSS offers equivalent perceived image quality with 4k TAA while giving extra performance, we'll have to agree native 4k rendering was overkill from the start and should have gone for lower internal rendering resolutions instead. (we were essentially wasting performance for no perceivable benefit)

Meanwhile AFAIK the current implementation of RT on BF5 runs denoising on compute cores, which means we're still waiting to see the true potential for tensor cores in consumer apps.

I'd like to see what DLSS 2X can do though, assuming the performance price is manageable.

 

[NTMBK]

Here you go: *SNIP*

Thanks! I can actually make out the details properly now. Yeah, some of the fine detail (e.g. on the column on the side of the building, lower right) has been lost on the DLSS version. Overall though, not bad for an upscaler.

 

[sontin]

Raytracing is a failure because of the performance loss. Image quality? Doesnt matter.
DLSS is a failure because of the image quality loss. Performance? Doesnt matter.

I like how one person can define a "failure" for two totally different techniques which are doing exactly the opposite.

BTW: DLSS looks much superior to TAA when you play FF15. But i guess playing games dont matter anymore, too.

 

[coercitiv]

Raytracing is a failure because of the performance loss. Image quality? Doesnt matter.
DLSS is a failure because of the image quality loss. Performance? Doesnt matter.


I like how one person can define a "failure" for two totally different techniques which are doing exactly the opposite.

Raytracing is a win because image quality is what matters most.
DLSS is a win because performance is what matters most.

I like how one person can define a "win" for two totally different techniques which are doing exactly the opposite.

BTW: DLSS looks much superior to TAA when you play FF15. But i guess playing games dont matter anymore, too.

DLSS looks arguably equivalent to 1800p TAA when watching the Infiltrator demo, which unlike FF15 comes with a proper TAA implementation.

This is a serious battle, in some situations I noticed fewer jagged edges with the upscaled 1800p image, and in some situations I noticed higher detail with the DLSS imagery. It’s an incredibly close comparison; where native 4K was that little bit sharper than DLSS, 1800p is a very good match for DLSS.

 

[PeterScott]

https://www.techpowerup.com/reviews/NVIDIA/DLSS_Super_Sampling_AA_Tested_Final_Fantasy/


Epic fail.

Article you quote pretty much says the opposite:

A major improvement that DLSS has over temporal anti-aliasing (TAA) is that motion vectors aren't used, a technique that leads to artifacts on TAA...

Compared to upscaled 1440p gaming on a 4K monitor, everything looks better, especially texts in the HUD are much crisper because they are rendered on top of the final DLSS image at the native 4K resolution (look at the player health bars in the bottom right). Upscaled 1440p does run faster than 4K+DLSS though, by another 33%.

Overall, I have to say that I'm pleased with how DLSS looks; the biggest caveat is game support, which is extremely lacking at this time. NVIDIA has confirmed several times that many game studios are working on DLSS support, so it looks like it'll just be a matter of time until it sees wider adoption. Personally, I see DLSS as an additional quality dial, to balance FPS vs quality, especially for weaker hardware. You can now increase settings on models and textures that would have previously driven FPS down below playable levels. DLSS will in turn bring the frame-rates back up.

Back when DLSS was a fresh topic, I downloaded HQ videos of FF and Infiltrator and compared them in detail.

IMO, in FF DLSS had an obvious edge over TAA, but it depended on frame examined, you can find frame to favor whichever argument you want to make.

In infiltrator, TAA had a slight edge. But it was a very close thing.

Overall I find DLSS less impressive than the marketing (of course). Initially it sounded like there would be one DLSS method that would be as good as SSAA at no overhead. Not its more like barely TAA, and there is a high quality DLSS that is yet unseen.

IMO it's a little underwhelming so far, but nowhere near an "epic fail".

 

[antihelten]

Based on the slides comparison. You have to zoom in quite a bit to spot the DLSS blur on some of them, something which does not happen in-game. antihelten's pics showing the blur are zoomed in past the games viewing point, and only on those selected pics. Other pics (2 & 5) clearly favor DLS, even when zoomed in. I could have done the same as antihelten and cherry picked those to make an argument of superior DLSS results. So mixed results as I see it (from our rather limited vantage points of examining pics).

Actually pic 2 doesn't really favour DLSS, since it is still clearly blurrier than the alternative (FXAA), it isn't nearly as pronounced as when comparing to the first picture though. One thing you still suffer from though is the fairly hefty over-sharpening artifacts you get with DLSS, which you don't get with the post-processing AA,

pic 5 is a 1440P image and one that runs 30% faster than DLSS, so you would bloody well hope that DLSS looks better here, otherwise it would obviously be a complete failure.

Also I actually already pointed out picture 5 in my original post (and noted that it looked worse than DLSS), so I don't know why you're accusing me of cherrypicking?

Thanks! I can actually make out the details properly now. Yeah, some of the fine detail (e.g. on the column on the side of the building, lower right) has been lost on the DLSS version. Overall though, not bad for an upscaler.

It's not just on the column, it's basically all over the place, wherever there is some kind of high frequency pattern, also as mentioned above DLSS also suffers from oversharpening artifacts (most noticeably the black halos around the vine leaves on the white column, the water pipe on the left and the inside of some of the windows on the bridge)

Not bad, but not really good either, just kinda ok, considering that TAA at 1800p apparently already achieves the same level of IQ and performance (based on the techspot article linked above).

All in all it looks like a lot of effort (on Nvidia's part) for apparently no improvement over TAA (although I would still like to see some comparisons with on screen motion).

 

[DisarmedDespot]

DLSS would've been much better recieved if it was a bonus feature on top of the card. Instead it and the raytracing cores are the only two things justifying the 20XX series price hikes, especially for the 2070 and 2080 which don't perform much better than the 1080 and 1080 Ti..