Nvidia DLSS 1 and 2 antialiasing discussion *spawn*

but I'm pretty sure at that distance they're not supposed to look like they do on the DLSS shot. When you get closer, sure, but not at that distance, regardless of how they look in TAA shot.
You are contradicting yourself, if they look like that up close, then there isn't any reason why they shouldn't look the same at a distance, if DLSS is correcting some of the LoD problems then this is more of than a welcome addition.
 
You are contradicting yourself, if they look like that up close, then there isn't any reason why they shouldn't look the same at a distance, if DLSS is correcting some of the LoD problems then this is more of than a welcome addition.
Err, what? Last time I checked things tend to lose details the further away they are in real life too, and in that particular scene with bright light relative to scene blasting on to said building there's no way they should be that defined.
 
NVIDIA DLSS 2.0 Behind the Scenes – How the Magic Happens
That is not to say there are no issues with TAA or CR. Since there are content changes when rendering games in real-time (the scenes are continuously dynamic, after all), naively assuming previous frames will be correct could easily lead to artifacts like ghosting or lagging. Usually, these problems are handled with heuristics based history rectification of the invalid samples from previous frames. That comes with other issues of its own, though, such as the reintroduction of temporal instability, blurriness, and Moiré patterns.

One of the most commonly used heuristics models is neighborhood clamping, which works by clamping samples from previous frames to a minimum and maximum of the neighboring current frame samples. This does strike a decent balance to avoid the shortcomings mentioned above. However, it cannot prevent a rather significant loss of detail as you can see below.


NVIDIA is solving this issue by exploiting the power of its supercomputer, which is trained offline on tens of thousands of extremely high-quality images.

Neural networks are simply much more suited to a task like this than handcrafted heuristics as they can find the optimal strategy to combine samples collected over multiple frames, delivering much higher quality reconstructions in the end result.

It's a data driven approach and one that allows DLSS 2.0 to successfully reconstruct even complex situations like those with the Moiré pattern. The image comparisons below are thoroughly impressive, surpassing even native images more often than not while doing a 4x upscaling from 540p to 1080p. None of this was possible with the previous DLSS model.
https://wccftech.com/nvidia-dlss-2-0-behind-the-scenes-how-the-magic-happens/
 
If DLSS doesn't need training per game any more, how long until we get to use it on old titles? Like, SD titles upressed to HD. Can DLSS cope with pixelised text?
DLSS 2.0 works great in "Deliver us the Moon". One picture is 1080 with TAA, the other is 4K with DLSS@Performance:

 
Err, what? Last time I checked things tend to lose details the further away they are in real life too,
Games usually have aggressive shadow LoDs that cut down shadows way before geometry, so no this isn't natural at all. The way I see it, this is the correct way to handle these windows, shadows don't magically disappear into thin air completely as in that TAA shot. I would rather have them, even if slightly accentuated than not having them at all.
 
Err, what? Last time I checked things tend to lose details the further away they are in real life too, and in that particular scene with bright light relative to scene blasting on to said building there's no way they should be that defined.
Not sure if this is fully comparable to what's happening on that Youngblood scene, but DLSS does seem to conjure up some detail that TAA cannot at a 100% resolution:


A notable dowside of DLSS here is that those white lines from the realms of super resolution are quite prone to breaking in motion.
 
A notable dowside of DLSS here is that those white lines from the realms of super resolution are quite prone to breaking in motion.
Isn't that true for either DLSS or TAA? The white lines are present in the first (TAA) and second (DLSS) images.
 
Interesting reading, but there's a real problem with how DLSS is being talked about at the moment. eg

"The image comparisons below are thoroughly impressive, surpassing even native images more often than not while doing a 4x upscaling from 540p to 1080p."
DLSS2_1-scaled.jpg


TAA is being applied and that's blurry. Why is DLSS constantly being compared to TAA processed images and no other AA modes, or even zero AA? I mean, 4K viewed from a reasonable distance isn't going to see an awful lot of jaggies.
 
Why is DLSS constantly being compared to TAA processed images and no other AA modes, or even zero AA? I mean, 4K viewed from a reasonable distance isn't going to see an awful lot of jaggies.
Aside from TAA I think the only other comparison currently made is against TAAU. Video started at 39:28.
 
Last edited by a moderator:
Interesting reading, but there's a real problem with how DLSS is being talked about at the moment. eg

"The image comparisons below are thoroughly impressive, surpassing even native images more often than not while doing a 4x upscaling from 540p to 1080p."
DLSS2_1-scaled.jpg


TAA is being applied and that's blurry. Why is DLSS constantly being compared to TAA processed images and no other AA modes, or even zero AA? I mean, 4K viewed from a reasonable distance isn't going to see an awful lot of jaggies.
Probably because most engines only have visuals which work due to TAA. UE4, idtech, frostbite, northlight, etc. have all designed many of their effects (ambient occlusion, reflections, transparency rendering, etc.) around stochastic accumulation and then cleaning that up with TAA. If you turn off TAA in these games, most of their effects works breaks. So there is really no alternative to TAA in many of these game engines, or if you choose to not use TAA (like turning it off in an ini file), then you have a dramatically worse looking game that no amount of MSAA or FXAA could save.
 
Isn't that true for either DLSS or TAA?
I guess so, but as TAA misses the lines altogether or draws them more faintly, the breaking in motion doesn't really stand out as much. Strong lines breaking in motion does.
The white lines are present in the first (TAA) and second (DLSS) images.
Just to clarify, the point of the post was to show that DLSS manages to draw detail that is missing in1440p TAA image but appears when we bump the resolution scale to 200% (2880p).

When I noticed these white lines as a difference between DLSS and TAA earlier, I was curious whether DLSS is drawing detail that shouldn't be there. Testing with an image closer to ground truth revealed that DLSS was indeed doing the better job here.
 
Probably because most engines only have visuals which work due to TAA. UE4, idtech, frostbite, northlight, etc. have all designed many of their effects (ambient occlusion, reflections, transparency rendering, etc.) around stochastic accumulation and then cleaning that up with TAA. If you turn off TAA in these games, most of their effects works breaks. So there is really no alternative to TAA in many of these game engines, or if you choose to not use TAA (like turning it off in an ini file), then you have a dramatically worse looking game that no amount of MSAA or FXAA could save.
How ironic that we push for higher resolution and then blur the results.

I very hope AMD can do something similar to DLSS and bring their own version to consoles. We'll get a very different experience combining NN with heuristics as a standard render target, and the results from NN are now really promising. Imagine 10+ TFs of rendering power only needing to target a 720p (or lower!) framebuffer; that'll be equivalent to 40 TFs rendering power at 1080p in terms of shader potential. Whereas for PC games, they are still targeting rendering higher resolution and just have the option for more frames when using DLSS.
 
It's tough ask for consoles lacking tensor core like operations. One would have to compare the tensor core flops/tops to what shaders on console can do. MS does have their 4x 8bit math instructions but that still falls quite ways short. But maybe they can figure out cheaper NN to do similar and optimize more as they have closed box instead of a lineup of hardware.

Maybe DNN's is the sauce that allows for next next generation or pro model without having to make hw that much more expensive. Add fancy neural network based software and NN acceleration hw.
 
I visited that same scene in Youngblood:


TSSAA 8X vs DLSS Quality vs No AA. No sharpening of any kind.

That no AA shot pretty clearly reveals that we are indeed supposed to be seeing those antennae and the framing around the windows.
I can agree on the antennae, but not on the window framings, clearly they're mostly gone at that distance with no AA?
 
I can agree on the antennae, but not on the window framings, clearly they're mostly gone at that distance with no AA?
They are partially drawn and as the image slightly moves as the character sways, a different part of the line gets revealed. Looks to me the resolution just isn't high enough to complete the lines. I assume if I could bump it to 200% the lines would be more complete. Sadly, exceeding a 100% resolution scale seems a bit tricky in Youngblood. Might give it another try later.

Even if my assumption was wrong here I'm hard pressed to see how DLSS isn't the clear winner here. The wall looks detailed and complete to me and the lines are pretty subtle and not over-emphasized.
 
They are partially drawn and as the image slightly moves as the character sways, a different part of the line gets revealed. Looks to me the resolution just isn't high enough to complete the lines. I assume if I could bump it to 200% the lines would be more complete. Sadly, exceeding a 100% resolution scale seems a bit tricky in Youngblood. Might give it another try later.

Even if my assumption was wrong here I'm hard pressed to see how DLSS isn't the clear winner here. The wall looks detailed and complete to me and the lines are pretty subtle and not over-emphasized.
perhaps if you use DSR there and set smoothing to 0% to get 4x resolution but nearest neighbour scale?
 
Back
Top