AMD FSR antialiasing discussion

  • Thread starter Deleted member 90741
  • Start date
This DLSS to FSR2.0 mod stuff is pretty cool for AMD owners!

FSR1.0 in Cyberpunk was pretty decent huh.. I'm surprised. The FSR2.0 mod looks better and sharper however. Still... pictures only tell a half of the story as we well know. Still image comparisons are for the most part useless (unless you're showing large movements of motion in them) as both techniques accumulate frames over time. Relatively still images will produce largely similar results.. whereas the real meaningful testing comes from how they compare under stressful situations.

FSR2.0 will undoubtedly improve from here, and that's pretty exciting.

Looking at still shots in God of War, I can pick out aspects of images where I like the FSR2.0 resolve slightly more than DLSS2.0.. and other aspects where it's the opposite. However, in motion, DLSS is still clearly superior. It's all bout that image stability.. and if you play at high output resolutions, it's pretty incredible how stable DLSS looks. In a world where there's so many graphical techniques being utilized to improve performance, which can cause lots of shimmering, breakup, and dithering... technologies like DLSS are game changers.

It's not perfect of course... none of them are, but considering what we're asking of them, it's stupidly impressive.
 
OlegSH said:
I wouldn't compare TAAU with the second gen reconstruction technologies, such as TSR, FSR 2.0, DLSS 2.0+ and XeSS at all.
The second gen technics are built around minimizing losses by getting rid of the lossy color clipping heuristic, optimizing resampling, preserving and resolving subpixel details.
FSR still relies upon color clipping, so some of TAAU shortcomings are still here. The pixel locking heuristic is to resolve thin details, which are otherwise removed by color clipping, but it relies upon the depth disocclusion mask and can introduce ghosting artifacts on the inner geometry parts. Judging by DF's analysis in GOW, pixels can be locked on parts that are not affected by the disocclusion mask (such as Kratos hands or on water), so there is nothing that can prevent ghosting from happenning on these parts expept for developer's provided masks. Overall that looks like a bunch of fragile hacks to me, these hacks can fix certain things, but break others, so games have to be carefully tuned on a case by case basis. Instead of building 1000 heuristics to handle different scenes in games, you can train a neural net. Understanding what is going on in frame, what parts of image to blend and what parts are better not touch is a key to quality and ease of integration and neural nets have proven to be the right tool for the job (vs handcrafted expert-based algorithms).
Click to expand...
I'm not saying they're not more advanced than usual forms of engine/in-game TAAU, just that they're still inherently the same thing. Using or not using neural nets makes no difference to it.
 
Kaotik said:
just that they're still inherently the same thing
Click to expand...
All temporal technics are the same, they all accumulate pixels over time, the same goes for checkerboarding or interlaced rendering, just input layouts are different.

Kaotik said:
Using or not using neural nets makes no difference to it.
Click to expand...
It makes a huge difference to quality that you can get out of the accumulation.
 
OlegSH said:
All temporal technics are the same, they all accumulate pixels over time, the same goes for checkerboarding or interlaced rendering, just input layouts are different.


It makes a huge difference to quality that you can get out of the accumulation.
Click to expand...
It can make a difference in quality, but it doesn't turn DLSS into something other than yet another form of temporal upscaling (with AA). That's where I draw my line anyway, spatial upscalers is another basket, while I would see for example checkerboarding being rather another form of interlacing than anything else (even if there's accumulation over more than just previous frame)
 
Kaotik said:
It can make a difference in quality, but it doesn't turn DLSS into something other than yet another form of temporal upscaling (with AA). That's where I draw my line anyway, spatial upscalers is another basket, while I would see for example checkerboarding being rather another form of interlacing than anything else (even if there's accumulation over more than just previous frame)
Click to expand...
"Interlacing" is also a temporal upscaling. If you consider any method in which a final frame is built from parts of a current and previous frame(s) then pretty much all of these are "the same".

But I feel that this is too generic of an approach the only purpose of which is to diminish the advantages of one temporal reconstruction technique over another.
 
Now FSR 2.0 released as open-source. FSR 1.0 NEVER got any improvements since it's source code release 1 year ago, But can we expect different situation for FSR 2.0?
Will it improve over time thanks to collective intelligence?
 
OlegSH said:
Agree. The point is all these technics produce results of different quality, so there is no sense in fixating on the "temporal upscaling" semantics when comparing them.
Click to expand...

Digital Foundry explained it before multiple times, they are totally different technologies with different results and in special different future-proofings. Intel and NV both seem to go the same routes whilest AMD isnt as of yet, but they suspect they will follow suit, for RT aswell.
 
DegustatoR said:
"Interlacing" is also a temporal upscaling. If you consider any method in which a final frame is built from parts of a current and previous frame(s) then pretty much all of these are "the same".

But I feel that this is too generic of an approach the only purpose of which is to diminish the advantages of one temporal reconstruction technique over another.
Click to expand...

On the contrary I think recognizing that all of these techniques are attempting to solve the same problem makes it much easier to discuss the advantages.

It’s only results that matter not the complexity of the implementation. Checkerboarding, spatial, temporal upscaling achieve varying levels of success in approximating or improving on a “native” image.

DLSS isn’t better because it’s fancier. It’s only better because the results are better.
 
trinibwoy said:
On the contrary I think recognizing that all of these techniques are attempting to solve the same problem makes it much easier to discuss the advantages.
Click to expand...
That's not what is happening though.
The approach from the problem being solved is valid since in such approach it is theoretically possible that some purely spatial technique could produce better results than a state of the art temporal one. It is also allowing us to compare all possible solutions, without dividing them into "bad" and "good" groups from the start.
But the approach from bundling all upscaling techniques into "technology groups" is pointless since there are no clear inherent benefits to any technology here. This applies not only to "spatial" vs "temporal" distinction btw but also to "algorithmic" vs "ML/AI-based" as well. Saying that "oh, these are all the same TAAU" completely misses the point - and actual reality in which said techniques may in fact be mixed on a user rated perception scale. A simple example here is FSR1 which is a really simple and purely spatial upscaling technique - which is generally considered to provide better (perf/qual) results than DLSS1 - a much more advanced technique with temporal and AI components.

There's also a little issue with calling DLSS2 "TAAU" btw - we don't really know what DLSS2 does to an image. The fact that it takes "TAAU-like" inputs doesn't mean that there are no purely spatial processing happening.
Same is true for FSR2 btw, which has RCAS in it, which is a purely spatial sharpener.
 
DegustatoR said:
That's not what is happening though.
The approach from the problem being solved is valid since in such approach it is theoretically possible that some purely spatial technique could produce better results than a state of the art temporal one. It is also allowing us to compare all possible solutions, without dividing them into "bad" and "good" groups from the start.
But the approach from bundling all upscaling techniques into "technology groups" is pointless since there are no clear inherent benefits to any technology here. This applies not only to "spatial" vs "temporal" distinction btw but also to "algorithmic" vs "ML/AI-based" as well. Saying that "oh, these are all the same TAAU" completely misses the point - and actual reality in which said techniques may in fact be mixed on a user rated perception scale. A simple example here is FSR1 which is a really simple and purely spatial upscaling technique - which is generally considered to provide better (perf/qual) results than DLSS1 - a much more advanced technique with temporal and AI components.

There's also a little issue with calling DLSS2 "TAAU" btw - we don't really know what DLSS2 does to an image. The fact that it takes "TAAU-like" inputs doesn't mean that there are no purely spatial processing happening.
Same is true for FSR2 btw, which has RCAS in it, which is a purely spatial sharpener.
Click to expand...

I’m not really following. If we evaluate all of these techniques on their merits then all those other details are just noise.

The debate on which techniques are inherently superior is interesting but ultimately pointless.
 
trinibwoy said:
I’m not really following. If we evaluate all of these techniques on their merits then all those other details are just noise.

The debate on which techniques are inherently superior is interesting but ultimately pointless.
Click to expand...
My point is that all upscaling techniques "compete" with each other in the same way as all antialiasing techniques compete with each other. There is zero reason to "bundle" them into "types" because this doesn't make anything clearer or better.
 
DegustatoR said:
My point is that all upscaling techniques "compete" with each other in the same way as all antialiasing techniques compete with each other. There is zero reason to "bundle" them into "types" because this doesn't make anything clearer or better.
Click to expand...

I agree.
 
FSR2.0 for Metro Exodus , Death Stranding , Marvel’s Guardians of the Galaxy :

videocardz.com

AMD FSR 2.0 mod reportedly also works in Death Stranding, Guardians of the Galaxy and Metro Exodus - VideoCardz.com

AMD FSR 2.0 mods expands into new games The PC modding community enables AMD temporal upscaling technology in more NVIDIA DLSS-enabled games. Redditors have been busy trying to figure out how to replace NVIDIA DLSS 2.0 with AMD FSR 2.0 in games such as Cyberpunk 2077 or Dying Light. As it turns...
videocardz.com videocardz.com


FSR2.0 for Red Dead Redemption 2 :

videocardz.com

Red Dead Redemption 2 gets unofficial FSR 2.0 mod - VideoCardz.com

Red Dead Redemption 2 with AMD FSR 2.0 support The PC modding community is enabling AMD FidelityFX Super Resolution 2.0 in more and more games. AMD FSR 2.0 Quality Mode vs Native at 1920×1080 resolution, Source: Github Following the first mod for Cyberpunk 2077 and its fork for Dying Light 2...
videocardz.com videocardz.com
 
You must log in or register to reply here.

Similar threads

Cyan
Microsoft's Auto SR: Universal AI upscaling discussion.
Replies
7
Views
2K
Janne Kylliö
J
B
DLSS versus FSR, again, neither being PSSR *spawn
2 3
Replies
41
Views
5K
homerdog
homerdog
Unknown Soldier
3DMark AMD FSR Feature Test
Replies
1
Views
1K
Putas
P
BRiT
Intel XeSS anti-aliasing discussion
9 10 11
Replies
201
Views
37K
Cyan
Cyan
M
Proposed upscaling technique - render vectors to pixels, convert to vectors, upscale and re-raster *spawn
2 3
Replies
43
Views
4K
JoeJ
J
Back
Top