Support NeoGAF

[Kazekage1981]

 

[01011001]

where is this image coming from btw? is that an official promo shot by AMD?

the real test will be motion clarity tho. standing still hugely benefits temporal up-sampling because it can use a ton of frames to get more information. in motion on the other hands this often really quickly breaks apart

 

[Fafalada]

It's interesting that on these two shots at least - both DLSS and FSR2 are outperforming native image handling of really thin features.
And it's DLSS that's oversharpening the image this time around. Will be interesting to see how it behaves in motion.

 

[Kazekage1981]

where is this image coming from btw? is that an official promo shot by AMD?

the real test will be motion clarity though. standing still hugely benefits temporal up-sampling because it can use a ton of frames to get more information. in motion on the other hands this often really quickly breaks apart

i took a prnt screen shot of that youtube video, pasted it on to windows paintbrush, saved it as a jpeg image, and scaled it down and posted it here.

 

[Fafalada]

in motion on the other hands this often really quickly breaks apart

I mean in motion they all kind of fall-apart, we just don't notice it much because average display at home has motion resolution that only resolves at about 25% of the actual panel pixels, so all the temporal methods play to display weaknesses... or is it strengths? Anyway you get the point.

That said the thin-feature resolving is way superior to any other temporal upscale here - it clearly resolves more detail than native image on those (and it makes sense, it's all explained in the talk on why/how), in a similar fashion DLSS does it.

But yes, want to see how motion stability holds up now.

 

[01011001]

ElAnalistaDeBits released a Ghostwire Tokyo video today for PC settings
and I made a small comparison pic with their shots using DLSS 2, FSR 1 and Epic's TSR

(very high res image, 2560x4103, be warned)

Spoiler: IMAGE

Temporal Super Resolution looks pretty damn good, that is the best we can realistically expect from FSR 2.0, how it actually looks natively on your screen and in motion is of course a different story, and that's where many reconstruction methods start to really show their weaknesses.

BUT, both DLSS and TSR look better than native at least in this comparison, as they both complete lines andfix weird flicker of effects that native exhibits.

 

[ethomaz]

which one of these do you think looks the best? I mixed them around and scaled it up 200%

sadly the og image has JPEG compression artifacts and is not really great quality... a high res png would have been better, but well...

Did you made another JPEG conversion or the IMGUR is converting again what you send them?
Anyway let's say the order is:

A B
C D

IMO from better to worst: A > C > D > B
C is the most oversharped.
D is some places blurry others a bit oversharped (less than C)
A basically has the best overall quality I believe but even so it is a bit blury in some places but there is no oversharpness.

 

[ethomaz]

BUT, both DLSS and TSR look better than native at least in this comparison, as they both complete lines andfix weird flicker of effects that native exhibits.

Is that native without any post AA applied?

 

[01011001]

Did you made another JPEG conversion or the IMGUR is converting again what you send them?
Anyway let's say the order is:

A B
C D

IMO from better to worst: A > C > D > B
C is the most oversharped.
D is some places blurry others a bit oversharped (less than C)
A basically has the best overall quality I believe but even so it is a bit blury in some places but there is no oversharpness.

here's the labeled version

Is that native without any post AA applied?

no it's with the game's TAA, which is always the only actually good AA solution for modern games due to high frequency detail.
of course you can use SSAA but you can't expect people to have the GPU overhead to actually do that at good framerates.

 

[Notabueno]

DLSS moved to a general algorithm as far as I know. Its not really using ML on the card per se, but a neural network, that's two different things. At least to my understanding.

Okay, so they're more or less the same now. DLSS used to have to be trained on a per-game basis but I didn't know that changed.

 

[Notabueno]

here's the labeled version

no it's with the game's TAA, which is always the only actually good AA solution for modern games due to high frequency detail.
of course you can use SSAA but you can't expect people to have the GPU overhead to actually do that at good framerates.

So FSR2 has more of a blurry image therefor less pixelated but with a loss color/range, while DLSS is crispier, but with more pixelated artifacts and a more dynamic color reconstruction.

Fuck, I can't choose between those two problems.

 

[Rudius]

So FSR2 has more of a blurry image therefor less pixelated but with a loss color/range, while DLSS is crispier, but with more pixelated artifacts and a more dynamic color reconstruction.

Fuck, I can't choose between those two problems.

The main difference seems to be the sharpening intensity.

 

[ethomaz]

here's the labeled version

no it's with the game's TAA, which is always the only actually good AA solution for modern games due to high frequency detail.
of course you can use SSAA but you can't expect people to have the GPU overhead to actually do that at good framerates.

Thanks.
That explain why I found it with less jaggies but a bit more blurry than normally native 4k is.

 

[ethomaz]

So FSR2 has more of a blurry image therefor less pixelated but with a loss color/range, while DLSS is crispier, but with more pixelated artifacts and a more dynamic color reconstruction.

Fuck, I can't choose between those two problems.

Maybe there is some settlings… I did not like both IQ results too.

PS. It is not a deal breaker and I can play games with that IQ… it just not of my liking at all.

 

[JeloSWE]

So FSR2 has more of a blurry image therefor less pixelated but with a loss color/range, while DLSS is crispier, but with more pixelated artifacts and a more dynamic color reconstruction.

Fuck, I can't choose between those two problems.

Both have tweakable sharpening settings. But they are not usually accessible to end users. I imagine DLSS could be made a bit softer and FSR2 a bit sharper.

 

[01011001]

So FSR2 has more of a blurry image therefor less pixelated but with a loss color/range, while DLSS is crispier, but with more pixelated artifacts and a more dynamic color reconstruction.

Fuck, I can't choose between those two problems.

always depends on the game's implementation
usuall DLSS isn't oversharpened, also in motion DLSS usually holds up better than TAA.

it's ball about the motion clarity with these.
DLSS' machine learning part is used to clean up the image from artifacts, which in return often means cleaner motion and less aliasing compared to other methods.

the lower the base resolution the more important this gets and the more other mothods break down usually.

at 4k with the highest preset I bet FSR2 and TSR can hold up pretty well compared to DLSS, but once you go lower it will be more obvious.

at a 4K target even FSR1.0 looks decent enough because there are just so many pixels to work with.

it really will get interesting once people get their hands on this and start going with more reasonable resolutions like 1080p and 1440p which are way more common among PC users than 4k

 

[Kenpachii]

1440p
motion

Then we will see how good fsr 2.0 is. FSR 1.0 was hot garbage for this, so curious to see how it holds up.