AMD FSR antialiasing discussion

[BacBeyond]

we do not delete the Kings Hunt images because it is rule number 1 on the eurogamer Website that we leave a geneaology of corrections. We do not amend the og article body or Text or images, but add to it and clarify it. That is how all corrections are done journalistically so people can understand where corrections came in context and we cannot be accused of Altering our content in a secretive way after the fact. Another example of that Praxis is in the World of academia where a paper with an error requiring clarificarion is not deleted from the Account, but an addendum or corrections or Update Is issued.

Great, so Add images of KingsHunt with DOF disabled instead of leaving up the highly misleading versions you currently show and move them to an "incorrectly made" section. Issue resolved. You are no longer misleading people.

Because that is one of the most misleading things I've ever read. You are claiming that TAAU has much better detail, while ignoring that a huge amount of detail is lost from the FSR and Native (which you conveniently don't include here) images from the DOF applied.

You are claiming that 1080p TAAU is superior to 4k Native.

There is Native 4k (left) and 1080p TAAU (right) from your own website.

Are you honestly saying that comparison isn't completely flawed because of the DOF change?

You absolutely need to include new images for Kings Hunt showing DOF disabled for Native and FSR to compare with TAAU DOF disabled.

Not to mention if you actually showed FPS differences it would have been obvious as DOF Disabled shows a sizable performance increase.

I used to highly regard your channel and site, but your refusal to fix this issue is throwing up huge red flags for your testing methodologies and bias.

 

[Blooper]

PCGH has this wrong as I posted earlier in the thread - Terminator does not turn on TAA U when you use the res scaler at all. You would have to enable it in the .ini separately or use unreal engine unlocker - that is why their comparison screens were actually with it off and just standard default UE4 upscale.
DLSS is completely different technology - its integration being different and more labour intensive is because it is doing something primarily different.

If you were to go to a siggraph conference or GDC, no developer would put FSR on the same band of computation as an image reconstruction technique in its totality because it is not doing that - yet the base of internet opinions is doing that. \FSRs weighted edge upscale would maybe be one part of a full image reconstruction technique that is much more complex in its entirety. Just thinking about FSRs characteristics I think can allow one to come to that conclusion. Why it is so much easier, simpler to integrate, costs so little in performance, and is slotted in after anti-aliasing? Why does AMD describe it as weighted edge upscaling with image sharpening? Instead of looking at those facets which allow one to conclude easily it is radically different than image reconstruction, the internet opinion is to want it to be image reconstruction.

If people keep wanting FSR to be something it is not, they are setting themselves up for the most intense disappointments or grand delusions that will need to be maintained through cognitive dissonance - as the first game that has DLSS in its latest iteration and FSR at the same time will be a bloodbath. It should not be a bloodbath rationally, as a rational person would say "FSR is not even doing the same thing at all, they are not real comparison points", yet prevailing opinion wants this to be AMDs DLSS competitor.

Not sure why you decided to reply to a post discussing adoption rates with "muh DLSS is a technically superior technique". Not that I disagree of course, but it's totally irrelevant to the discussion.

As I was talking about in that post, I do not see any reason to believe that adoption rate of FSR will be poor, nor that it has started out poor. I provided a comparison point to DLSS because it's really the only other upscaling/image reconstruction technique that has seen mass adoption on the desktop. TAAU and CBR are both used fairly often on the consoles, but to a much, much lower degree on the desktop. If FSR exhibits a similar or even faster adoption rate to the only one of these 3 that has seen significant adoption on the desktop within just the first few days since launch, then I believe my argument still holds and is consistent.

 

[OlegSH]

So you are claiming the post is meaningless and that DOF isn't broken when TAAU is enabled?

That's a slight DOF convolution.

That's not your average DOF, this DOF is a few pixel size convolutional kernel that works near the camera so you have to move the camera to a certain distance or angle to the character so that it would take any effect only on the character, but the rest of the image is not touched by the effect as it should be and is still the same for the slight DOF on/off.

Even if the camera distance to the character in DF video is enough for the effect to take place (which I am not sure of), there is the whole scene behind the character that is not affected with the slight DOF, so you can still pay attention to these larger parts of the scene which are way worse with spatial upscaling anyway.

Second, this slight DOF has nothing to do with background DOF that works perfectly fine in all modes in DF video, so their comparisons are still perfectly valid.

I am really getting tired of these pointless discussions on FSR.

With FSR, you are getting exactly what you're paying for, that's a spatial upscaler, so there is no way it can avoid vast detail losses, what are you even arguing with here?

You can't get more details out of a thin air, that's not how the things work. By rendering 59% of pixels and resampling them to Native resolution with the Ultra Quality FSR you are getting the 59% of texture details assuming 1 to 1 texel to pixel density. The rest is a cheapo trick with sharpenning image to a point where it would look similar to native (of cause you can always do the same for native resolution).

If you want a similar quality to FSR just slap a bit of sharpenning in sweetfx, driver's CP, freestyle, reshade, etc on top of any ingame upscaling and here we go - https://imgsli.com/NTg5MzU/0/1

You can only improve the appearence of geometry edges, which is the only benefit and innovation of FSR (kind of since there is quite a bit of research in the edge aware spatial upscaling field), but that's a minor portion of pixels on screen anyway, so one would likely not even notice this (I did some AB testing and nobody spotted that).

If you want to go beyond this simple math there is no other way around rather than accumulating pixels across time which is exactly what TAAU, checkerboarding and all other temporal solutions do.

 

[Rootax]

PCGH has this wrong as I posted earlier in the thread - Terminator does not turn on TAA U when you use the res scaler at all. You would have to enable it in the .ini separately or use unreal engine unlocker - that is why their comparison screens were actually with it off and just standard default UE4 upscale.
DLSS is completely different technology - its integration being different and more labour intensive is because it is doing something primarily different.

If you were to go to a siggraph conference or GDC, no developer would put FSR on the same band of computation as an image reconstruction technique in its totality because it is not doing that - yet the base of internet opinions is doing that. \FSRs weighted edge upscale would maybe be one part of a full image reconstruction technique that is much more complex in its entirety. Just thinking about FSRs characteristics I think can allow one to come to that conclusion. Why it is so much easier, simpler to integrate, costs so little in performance, and is slotted in after anti-aliasing? Why does AMD describe it as weighted edge upscaling with image sharpening? Instead of looking at those facets which allow one to conclude easily it is radically different than image reconstruction, the internet opinion is to want it to be image reconstruction.

If people keep wanting FSR to be something it is not, they are setting themselves up for the most intense disappointments or grand delusions that will need to be maintained through cognitive dissonance - as the first game that has DLSS in its latest iteration and FSR at the same time will be a bloodbath. It should not be a bloodbath rationally, as a rational person would say "FSR is not even doing the same thing at all, they are not real comparison points", yet prevailing opinion wants this to be AMDs DLSS competitor.

- The "internet opinion" doesn't care if it is image reconstruction or upsampling or whatever. It's a boost in performances for a reasonable PQ cost. That's it.

- And of course it's AMDs DLSS competitor, as it's their way to boost FPS with a reasonable PQ cost. When RDNA2 launched, a part of the aftermatch was "ok, so, no DLSS alternative/competitor ?". And the result is FSR. They took a different route and technical approach of course, but the goal is the same...

 

[pharma]

At least on DOTA 2, FSR competitor is a sharpening filter ...

Chapters
00:00 Benchmark Numbers and FPS Difference, on GTX 1660 Ti and RX 570 4G
00:46 FSR and Quality Comparison, Can you tell the difference?
03:32 In-depth 4K 50%, 75% and 100% Comparison.
08:14 Thoughts on AMD FSR in Dota 2 and ending driver support for older cards.
1:55 Question of the Day, where can I get Windows 11?

 

[OlegSH]

Add images of KingsHunt with DOF disabled instead of leaving up the highly misleading versions

Highly misleading? Are you blind?

https://forum.beyond3d.com/threads/amd-fsr-antialiasing-discussion.62394/page-32#post-2212083
Now that that reddit guy has been finally able to enable TAAU (thanks to god for that!):
Performance FSR DOF Force Disabled:

https://imgur.com/t5kBOOI

TAAU 50% (Performance) DOF Force Disabled:

https://imgur.com/fRRscfe

Don't you really see the very same difference as in DF's video? Look at the huge detail loss on her skirt then.

Now that reddit guy has to choose less finicky scene and camera angles.
I understand, he has to prove his point, but staring at the wall with minimum amount of details and under her skirt is not representetive of neither real gameplay or image quality evaluation.
Then he has to enable RCAS so that the same cheap trick with sharpenning is applied for both methods.
And finally he has to adjust texture mip bias so that TAAU can work with Native res textures as it should be.
This would be representetive of the real TAAU implementation in a real game.
Then we can talk!

PS and god those JPGs are awful, he should upload PNGs to imgsli.

 

[orangpelupa]

on

At least on DOTA 2, FSR competitor is a sharpening filter ...

1080p screen the differences are miniscule (the sharpen filter looks oversharpen in some scenes tho)

 

[PSman1700]

For 99.9% of PC gamers, FSR will indeed be seen as a competitor to DLSS, regardless of the underlying technology.

- The "internet opinion" doesn't care if it is image reconstruction or upsampling or whatever. It's a boost in performances for a reasonable PQ cost. That's it.

- And of course it's AMDs DLSS competitor, as it's their way to boost FPS with a reasonable PQ cost. When RDNA2 launched, a part of the aftermatch was "ok, so, no DLSS alternative/competitor ?". And the result is FSR. They took a different route and technical approach of course, but the goal is the same...

They aint competitors since their playing in totally different leuges regarding both performance and image quality.
AMD currently doesnt have a real answer to DLSS, they will, but not this gen (RDNA2). Also, FSR will most likely see limited to no use at all in the console space, you guys are fighting the wrong battle.

 

[Bondrewd]

they will, but not this gen (RDNA2)

Did you just >imply miSIMD on client?
Lol NOPE.

 

[DavidGraham]

FSR is:

1 - Free;
2 - Working on all hardware, not just a subset of the portfolio of one IHV;
3 - Brings a substantial performance uplift on a number of resolutions without discernable image degradation to most on Ultra + high res;
4 - Super useful for low power hardware where potato mode settings aren't enough to run the game;
5 - Free IQ upgrade with no performance penalty when used with DSR/VSR.

So is TAAU, it even provides better details than FSR.

 

[DegustatoR]

They aint competitors since their playing in totally different leuges regarding both performance and image quality.

They are competitors since you can substitute DLSS with FSR and get the same result which you were aiming for - more performance. Performance gains are comparable. The fact that it won't be as good in IQ doesn't mean anything.
But I wouldn't expect many games to do such substitution due to the advantages of DLSS. So the real "competitor" to FSR will be TAAU.
At least until FSR 2.0.

 

[Rootax]

They aint competitors since their playing in totally different leuges regarding both performance and image quality.
AMD currently doesnt have a real answer to DLSS, they will, but not this gen (RDNA2). Also, FSR will most likely see limited to no use at all in the console space, you guys are fighting the wrong battle.

The performance boost is here...

For the PQ, you're right, but ultra quality is decent imo.

But you can't say they're not competitor because one solution is better than the other.

FSR is amd answer to dlss. And it's inferior for now, and it's ok to be.

 

[PSman1700]

They are competitors since you can substitute DLSS with FSR and get the same result which you were aiming for - more performance. Performance gains are comparable. The fact that it won't be as good in IQ doesn't mean anything.
But I wouldn't expect many games to do such substitution due to the advantages of DLSS. So the real "competitor" to FSR will be TAAU.
At least until FSR 2.0.

And thats why i suspect that if AMD wants to compete with something like DLSS, we wont be seeing that before RDNA3+ architectures.

But you can't say they're not competitor because one solution is better than the other.

FSR is amd answer to dlss. And it's inferior for now, and it's ok to be.

Its a substitute for TAAU-like upscalers since TAAU cant be used just on about any game engine whereas FSR can. FSR without the use of AI/ML and/or hardware acceleration wont ever stand a chance to something like DLSS, yes FSR can be improved but so will DLSS, the comparison is the wrong one, like Alex wrote before.

FSR seems to be a nice addition to the pc gaming space, especially on older/lower end hardware configurations. A competitor to DLSS? Nah, if so, a very bad one if where going to compare those.
For the sake of AMD, lets hope they dont put it as one.

And again, FSR is and will be a pc centric solution to begin with. You guys are barking up the wrong tree.

 

[trinibwoy]

FSR isn’t a substitute for temporal upscalers in an academic sense but it’s 100% a competitor from an end user’s point of view.

I agree DF should just post new images with DOF disabled. That issue is just a distraction from the fact that TAAU produces a much better result even with DOF properly disabled in the FSR shot. There is no controversy there.

The results so far seem to indicate that FSR does well with a relatively high input resolution at preserving texture detail and enhancing edge detail in some scenes. FSR also results in an obvious loss of fine detail in other scenes (see Terminator). As OlegSH points out this is expected because FSR has significantly less information to work with and is not attempting to reconstruct fine details.

Naive display scaling: Blurs the lower resolution frame while upscaling. Does not recover missing details. Blurring can be addressed somewhat with sharpening.

FSR: Avoids blurring the lower resolution frame while upscaling given a high enough input resolution. Not particularly useful at lower input resolutions. Does not recover missing details. Applies some sharpening.

TAAU/DLSS: Avoids blurring the lower resolution frame while upscaling. Gives useful results down to 50% input resolution. Adds missing detail from the lower resolution frame using previous frame history. Applies some sharpening.

I think it’s that simple.

 

[DegustatoR]

And thats why i suspect that if AMD wants to compete with something like DLSS, we wont be seeing that before RDNA3+ architectures

They don't want to compete. They want an upscaling solution which will work on their products. Such solution don't have to be FSR and where FSR is right now - it may not even be FSR which would be the best such solution.

 

[PSman1700]

They don't want to compete.

Well, theres the thing then, its a nice addition to the pc gaming space where this might see some use where theres no taau available in an engine (nom ue games etc).

it may not even be FSR which would be the best such solution.

We can asume amd will come with a reconstruction tech using AI/ML, accelerated on a hw block that will compete with NV’s solution.

The differences between FSR upscaling and dlss reconstruction using hw acceleration/ML in games are hust too large to call them competitors.

The whole underlying tech is different aswell. FSR tries to reach the same goals but doesnt compete.
In that vein we could be calling FXAA/SSAA or simply lowering the resolution and sitting further away a ’competitor’ aswell. Hech, lowering some settings could be a nice ’competitor’. All are solutions to trade performance for lower IQ.

 

[Rootax]

So now when a company propose an inferior solution, we will say that they just don't want to compete ?

 

[iroboto]

NA waking up to this thread:

Does anyone have a package of comparison PNGs I can run my IQ tools on?

 

[troyan]

The results so far seem to indicate that FSR does well with a relatively high input resolution at preserving texture detail and enhancing edge detail in some scenes. FSR also results in an obvious loss of fine detail in other scenes (see Terminator). As OlegSH points out this is expected because FSR has significantly less information to work with and is not attempting to reconstruct fine details.

Naive display scaling: Blurs the lower resolution frame while upscaling. Does not recover missing details. Blurring can be addressed somewhat with sharpening.

FSR: Avoids blurring the lower resolution frame while upscaling given a high enough input resolution. Not particularly useful at lower input resolutions. Does not recover missing details. Applies some sharpening.

TAAU/DLSS: Avoids blurring the lower resolution frame while upscaling. Gives useful results down to 50% input resolution. Adds missing detail from the lower resolution frame using previous frame history. Applies some sharpening.

I think it’s that simple.

FSR doesnt avoid blurring. In Kingshunt it is slightly sharper than the standard in engine upscaling: https://imgsli.com/NTkwNjQ/0/1

 

[cho]

NA waking up to this thread:
Does anyone have a package of comparison PNGs I can run my IQ tools on?

https://drive.google.com/drive/folders/13QB0f6SCO-g1K5-ATvdwHQWnJcL37BVI?usp=sharing