Monday, February 18th 2019

NVIDIA: Image Quality for DLSS in Metro Exodus to Be Improved in Further Updates, and the Nature of the Beast

NVIDIA, in a blog post/Q&A on its DLSS technology, promised implementation and image quality improvements on its Metro Exodus rendition of the technology. If you'll remember, AMD recently vouched for other, non-proprietary ways of achieving desired quality of AA technology across resolutions such as TAA and SMAA, saying that DLSS introduces "(...) image artefacts caused by the upscaling and harsh sharpening." NVIDIA in its blog post has dissected DLSS in its implementation, also clarifying some lingering questions on the technology and its resolution limitations that some us here at TPU had already wondered about.

The blog post describes some of the limitations in DLSS technology, and why exactly image quality issues might be popping out here and there in titles. As we knew from NVIDIA's initial RTX press briefing, DLSS basically works on top of an NVIDIA neural network. Titled the NGX, it processes millions of frames from a single game at varying resolutions, with DLSS, and compares it to a given "ground truth image" - the highest quality possible output sans any shenanigans, generated from just pure raw processing power. The objective is to train the network towards generating this image without the performance cost. This DLSS model is then made available for NVIDIA's client to download and to be run at your local RTX graphics card level, which is why DLSS image quality can be improved with time. And it also helps explain why closed implementations of the technology, such as 3D Mark's Port Royal benchmark, show such incredible image quality scenarios compared to, say, Metro Exodus - there is a very, very limited number of frames that the neural network needs to process towards achieving the best image quality.
Forumites: This is an Editorial

The nature of DLSS means that the network needs to be trained for every conceivable resolution, since different rendering resolutions will require different processing for the image to resemble the ground truth we're looking for. This is the reason for Metro Exodus' limits in DLSS - it's likely that NVIDIA didn't actually choose not to enable it at 1080p with RTX off, it was just a case of the rendering time on its NGX cluster not being enough, in time for launch, to cover all of the most popular resolutions, with or without RTX, across its three available settings. So NVIDIA mated both settings, to allow the greatest image quality and performance improvements for those gamers that want to use RTX effects, and didn't train the network for non-RTX scenarios.
This brings with it a whole lot of questions - how long exactly does NVIDIA's neural network take to train an entire game's worth of DLSS integration? With linear titles, this is likely a great technology - but apply this to an open-world setting (oh hey, like Metro Exodus) and this seems like an incredibly daunting task. NVIDIA had this to say on its blog post:
For Metro Exodus, we've got an update coming that improves DLSS sharpness and overall image quality across all resolutions that didn't make it into day of launch. We're also training DLSS on a larger cross section of the game, and once these updates are ready you will see another increase in quality. Lastly, we are looking into a few other reported issues, such as with HDR, and will update as soon as we have fixes.
So this not only speaks to NVIDIA recognizing that DLSS image quality isn't at the level it's supposed to be (which implies it can actually degrade image quality, giving further credence to AMD's remarks on the matter), but also confirms that they're constantly working on improving DLSS' performance and image quality - and more interestingly, that this is something they can always change, server-side. I'd question the sustainability of DLSS' usage, though; the number of DLSS-enabled games is low enough as it is - and yet NVIDIA seems to be having difficuly in keeping up even when it comes to AAA releases. Imagine if DLSS picked up like NVIDIA would like it (or would they?) and expand to most of the launched games. Looking at what we know, I don't even think that scenario of support would be possible - NVIDIA's neural network would be bottlenecked with all the processing time required for these games, their different rendering resolutions and RTX settings.
DLSS really is a very interesting technology that empowers the RTX graphics card of every user with the power of the cloud, as NVIDIA said it did. However, there are certainly some quirks that require more processing time than they've been given, and there are limits to how much processing power NVIDIA can and will dedicate to each title. That the network needs to be trained again and again and again for every new title out there bodes well for a controlled, NVIDIA-fed games development environment, but that's not the real world - especially not with an AMD-led console market. I'd question DLSS' longevity and success on these factors alone, whilst praising its technology and forward-thinking design immensely.
Source: NVIDIA
Add your own comment

112 Comments on NVIDIA: Image Quality for DLSS in Metro Exodus to Be Improved in Further Updates, and the Nature of the Beast

#26
EarthDog
They wouldnt and dont... it's too variable comparatively. Much much smaller dataset to work with.
Posted on Reply
#27
Xzibit
EarthDogThey wouldnt and dont... it's too variable comparatively. Much much smaller dataset to work with.
If the scene is similarly on rails, In-game no player control cut scene or a transitional dialog NPCs scene which there is no player interaction. It would there for not benefit ?
Posted on Reply
#28
EarthDog
I really cant answer with any certainty. But outside of cut scenes, games aren't on rails and are rendering millions of DIFFERENT fps in a session. Hence the difference. A benchmark gets to learn the EXACT same scene and since it is never different it can optimize.

I'd be interested to hear more details on the neural network and learning. Can't say I buy the opinion of the editorial (not enough horsepower to catch up and optimize games), but I can sure see why that is his conclusion.
Posted on Reply
#29
Crackong
So, DLSS works best on "DEMO" or "On rail benchmarks" but epic failed in the real game compared to simple downscale method ?
Posted on Reply
#30
Xzibit
EarthDogI really cant answer with any certainty. But outside of cut scenes, games aren't on rails and are rendering millions of DIFFERENT fps in a session. Hence the difference. A benchmark gets to learn the EXACT same scene and since it is never different it can optimize.

I'd be interested to hear more details on the neural network and learning. Can't say I buy the opinion of the editorial (not enough horsepower to catch up and optimize games), but I can sure see why that is his conclusion.
Well Nvidia did say
NvidiaDLSS uses fewer input samples than traditional techniques such as TAA
If it increases those samples it would need more resources.
Posted on Reply
#31
EarthDog
They also said with more time, it learns better... so not sure how to take that.
Posted on Reply
#32
hat
Enthusiast
I'm not even sure what to make of this. We now have an "AA" (upscaling) method that requires "training" on some cluster somewhere, on a per game, per resolution basis. It seems overly complicated... like flying a spaceship to work, which is 3 blocks away from your house.
Posted on Reply
#33
Crackong
EarthDogThey also said with more time, it learns better... so not sure how to take that.
So we should play the game maybe a year after the release for optimal DLSS performance ?
Is this why Metro choose to release on steam at 2020 ?
Posted on Reply
#34
EarthDog
CrackongSo we should play the game maybe a year after the release for optimal DLSS performance ?
Is this why Metro choose to release on steam at 2020 ?
I'll assume you are serious here...

No. Play it as is and if the IQ losses are too great... dont use it. See an update that is supposed to improve it... check again. Dont make it complicated. ;)
Posted on Reply
#35
Xzibit
hatI'm not even sure what to make of this. We now have an "AA" (upscaling) method that requires "training" on some cluster somewhere, on a per game, per resolution basis. It seems overly complicated... like flying a spaceship to work, which is 3 blocks away from your house.
Well if you take Nvidias word. Its being trained on SS images of the current content. The improved A.I. algo is then being applied to app/drivers thus the algo is just being accelerated on the tensors GPU side.
Posted on Reply
#36
voltage
so weird how negative people are. they jump on the hate bandwagon so quickly with everything, including new hardware designs. If you never support something new in hardware, that might not me super useful now, but could be down the road you'll never get anything new from hardware makers that MIGHT bring good change for game designers. honestly, the world has gone nuts with HATE, and sad to see that seep into an industry that I considered a hobby for so very long. This mentality will over time stifle and or slow down hardware development, and all you'll be getting is faster cards with no attempt made a something new. The worst one-sided hate towards nvidia I have seen yet, is the latest from Hardware Unboxed. now ill most likely get no understanding for what I just typed, except more hate. really sad to see this going on
Posted on Reply
#37
moproblems99
XzibitSS images
Wouldn't the images be black and white? I didn't think they had color back then.
voltageIf you never support something new in hardware, that might not me super useful now
The idea is that those not super useful features are not the selling point but rather a bonus. Anyway, I think they amount of performance they managed to squeeze out of them is more impressive than the two features.
Posted on Reply
#38
Patriot
Alright guys...
DLSS is using pre-trained models and then running Inferencing on the tensor cores.
This is a far lighter workload than training. Training is done on DGX clusters, driver gets the new model.
Same thing self driving cars do, its pattern recognition based on pre-trained models... with months of compute time to back it.

That said... it may get better it may not, neural nets are not a one size fits all... one model may be good for one area of the map and terrible for another.
I may have facepalmed when 3dmark and other completely fucking linear tests are used as what DLSS can be.
It simply will never get there, just by the nature of the technology being used, sure it might get a skosh better... but not unicorns shitting rainbows 3dmark lies.
Posted on Reply
#39
Crackong
voltageso weird how negative people are.
People paid for the promises that Jensen made and it clearly doesn't work.
People get mad.
Is that so weird?
moproblems99The idea is that those not super useful features are not the selling point but rather a bonus.
You do know Jensen spent over 20 minutes on RTX release event On Stage talking about this "not the selling point" feature, right ?
Posted on Reply
#40
PrEzi
DLSS IS Garbage.
Look at the video test/review and comparison done by Hardware Unboxed --- you are better off with your own upscaling - textures more crisp, more details, and same/similar framerates to DLSS.
www.youtube.com/watch?v=3DOGA2_GETQ
Posted on Reply
#41
Tsukiyomi91
Anyone here tested without turning on any level of anti-aliasing but leaves DLSS enabled? Maybe that's how it supposed to work alone? Since it's meant to replace anti-aliasing without giving more performance hit while maintaining decent image quality?
Posted on Reply
#42
Crackong
Tsukiyomi91Anyone here tested without turning on any level of anti-aliasing but leaves DLSS enabled? Maybe that's how it supposed to work alone? Since it's meant to replace anti-aliasing without giving more performance hit while maintaining decent image quality?
Jensen at RTX launch event deliberately compare 4K TAA to 4K DLSS.
They are supposed to compete.
Posted on Reply
#43
Tsukiyomi91
That demo can't be used as a benchmark since it's a demo... I'm talking about real world comparison.
Posted on Reply
#44
laszlo
basically the AI can't process in real time each frame and won't be able to do it in the near future also as it need a lot a hp and time for learning...

nv you better leave out AI from gaming products as some gamers need max fps in max details instantly ...
Posted on Reply
#45
Tsukiyomi91
Well, you can turn off DLSS once you enable DXR/RTX feature on...
Posted on Reply
#46
Crackong
Tsukiyomi91That demo can't be used as a benchmark since it's a demo... I'm talking about real world comparison.
BFV DLSS is real world comparison.
And DLSS clearly lost to good old downscaling method.

And btw since it the "The Jensen" himself shown TAA vs DLSS On Stage with a DEMO.
I think it is fair.
And now we knew DLSS only works on DEMO or railed Benchmarks sections because they are fixed so the AI can train on them.

I assume the only type of "Real Game" that DLSS work best are old laser gun arcade shooters such as House of the DEAD.
Posted on Reply
#47
Tsukiyomi91
seems it is... Still, at least we know the Tensor cores are getting "trained" with "patches". But, it is still nowhere as we are expecting them to perform unless there's some sort of SDK to leverage those cores that are seemingly doing nothing.
Posted on Reply
#48
Crackong
Tsukiyomi91seems it is... Still, at least we know the Tensor cores are getting "trained" with "patches". But, it is still nowhere as we are expecting them to perform unless there's some sort of SDK to leverage those cores that are seemingly doing nothing.
If DLSS is just another SLI , user have to wait for SLI profiles and "Hopefully" it works. It is a waste of time and money.
Also speaking about time,
Nobody knows how long it would take for " Jensen's AI" to train for ONE game.
In this case for BFV, BFV released back in Nov 2018.
It Feb 2019 right now and his AI cannot produce better image than a simple downscaling method.
By the time the AI is capable to do that, the game maybe obsolete...........
Posted on Reply
#49
Ferrum Master
The worst thing... what if I do not play those few AAA games that have this tech? Waste of silicon? What if a game has the feature, but someone didn't coin enough to dedicate almighty nvidia server resources to tailor it for the exact game, and then the priorities, who decides which one is the priority now? BFV or Metro or whatever? If the game has mods, skins, reshade... all that thing goes to the toilet instantly as the pre AI training isn't aware of custom things.

I used nvidia 980ti and 1080ti... I am not fond of investing in things I don't use... We do not buy Quadro cards exactly for the same reason. Raytracing is nice, but this? No thanks...
Posted on Reply
#50
medi01
tvamosDon't know why TPU keeps using this image from Port Royal bench, when all images from real games show DLSS to be WAY more blurry.
To make it look better.
Why would they want it to look better than it is, is another question: perhaps stock? Or that wonderful NDA? Or basic green brains syncrome?

Or, perhaps, just lazy copypasta from green infected place.
Posted on Reply
Add your own comment
Nov 19th, 2026 20:56 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts