Fidelity Super Resolution in 2021

[james.miller]

You can thats what MSE is for, see the wiki article I linked too.

If, IF you were using MSE to test jpeg compression then your argument would be valid. But you arent, you're using it to test the rendering quality of FSR and DLSS against native and ignoring that lossy compression affects the image quality of the saved screenshots used to do the comparison. It's an uncontrolled variable therefor it's an invalid test.

 

[Rroff]

You'd need a scene which has no animated elements at all or the ability to freeze a frame and render multiple times to really compare accurately. Things like post-processing sharpening or where processing might produce a clearer output than native i.e. text are also going to be difficult to accurately gauge the impact of.

 

[muon]

The author was not wrong, you were. The authors program gets different values but the conclusion is the same. DLSS is almost the native image.

The flaw is the camera angle though.

Most other random differences like compression artefacts or stochastic elements would likely cancel out over an entire scene.

What a joker, yesmanning for zx128k but not once looking at the material til now.

But no one provided a valid argument until now. The screenshot of the comment thread posted didn't discredit anything.

 

[zx128k]

The flaw is the camera angle though.

The DLSS image SSIM is 0.9530 its almost completely the same as the native sample. 1 = the same. The difference between DLSS and the native sample is basically zero.

The MSE is 66.8661 which means the DLSS sample is basically very close to the native sample.

 

[muon]

The DLSS image SSIM is 0.9530 its almost completely the same as the native sample. 1 = the same. The difference between DLSS and the native sample is basically zero.

But that could be because the camera angle is closest aligned.

The SSIM thing is just a misleading criticism. As you've discovered it says the same thing as MSE.

 

[zx128k]

But that could be because the camera angle is closest aligned.

No it cant mean that the camera is closest aligned because we are ignoring FSR. It menas that the camera alignment between the DLSS sample and the native sample are close and that both images are very close to being the same. DLSS has near native quality, going by these samples. The DLSS sample has a SSIM close to being 1 with native sample. MSE is very low.

There does not appear to be a major issue between the DLSS and native sample. Looking at the FSR sample it does look off. The angle on the fence at the bottem is different. So I dont believe the FSR sample should be valid, the angle is off. The FSR sample gives much higher values for MSE and is lower for the SSIM value.

 

[Hotwired]

But no one provided a valid argument until now. The screenshot of the comment thread posted didn't discredit anything.

A joker once more, I posted the link to the reddit thread in the very first post which you claimed to have seen me post "10 times".

The fact you didn't care to read it doesn't mean it wasn't there.

 

[muon]

A joker once more, I posted the link to the reddit thread in the very first post which you claimed to have seen me post "10 times".

The fact you didn't care to read it doesn't mean it wasn't there.

You posted a screenshot and circled the elements which you believe supported what you were saying. It didn't.

Either way, I do agree now that the comparison is meaningless.

 

[Hotwired]

You posted a screenshot and circled the elements which you believe supported what you were saying.

I specially screenshotted the author making the initial post and later agreeing it was flawed. Which was the purpose of the circles being around their name not any post. Accompanied by links to the thread for context.

 

[muon]

I specially screenshotted the author making the initial post and later agreeing it was flawed. Which was the purpose of the circles being around their name not any post. Accompanied by links to the thread for context.

Okay fine. No point disagreeing on this point. You are correct the comparison is not remotely reliable.

 

[Rroff]

I specially screenshotted the author making the initial post and later agreeing it was flawed. Which was the purpose of the circles being around their name not any post. Accompanied by links to the thread for context.

You seem to be equating the author acknowledging it is flawed with it being useless - which it may or may not be. That it can be improved on doesn't mean the conclusion is necessarily different.

 

[Hotwired]

You seem to be equating the author acknowledging it is flawed with it being useless - which it may or may not be. That it can be improved on doesn't mean the conclusion is necessarily different.

The entire farce of the last few pages was about it not being the evidence it was being presented as which was just the initial post.

I think there's a good chance DLSS can be viewed as more accurate than FSR but incorrect work being repeatedly posted (here) as if it was quality isn't the way.

 

[zx128k]

I am running some images from here https://www.reddit.com/r/nvidia/comments/om5po7/dlss_vs_fsr_comparison_in_marvels_avengers/
DLSS Quality 4k vs native
MSE = 673.7376
SSIM = 0.7706

FSR ultra quality 4k vs native
MSE = 1.0349e+03
SSIM = 0.7092

DLSS Quality 1440p vs native
MSE = 750.0957
SSIM = 0.7266

FSR ultra quality 1440p vs native
MSE = 207.0987
SSIM = 0.8850

DLSS Quality 1080p vs native
MSE = 958.9497
SSIM = 0.6813

FSR ultra quality 1080p vs native
MSE = 355.0012
SSIM = 0.8311

I dont think these images can be compared. Just look at FSR ultra quality 4k vs native that MSE value.

 

[Rroff]

I think there's a good chance DLSS can be viewed as more accurate than FSR but incorrect work being repeatedly posted as if it was quality isn't the way.

There seems to be very different results between games as well - DLSS in Necromunda has a boat load of issues with reconstructed details, while both FSR and DLSS just look nasty if you reduce resolution and/or upscaling quality in that game - FSR especially starts to amplify image instability and/or overly sharpen jagged edges, etc. While Marvel's Avengers seems to have a much better implementation.

Too many comparisons as well with a base of 70-100+FPS into 120-200FPS which doesn't really show how well either of them can really do when stuff hits the wall - though nice if you want 120Hz gaming at 4K. Things get much more interesting where you are running around 40 FPS base and turning that into useable FPS.

 

[nvidiamd]

Can't believe what i read.
Especially in this game it is said that DLSS is making the image "better than native". That will mean by definition that any comparison with the native image will show errors in the DLSS screenshot. Heck FSR can be 99.99% as accurate as the native image and DLSS only 80% as accurate and DLSS image can still look better if the native image is bad.
So what is the purpose of such comparisons? Take the images with that bridge for example, assuming they are real. Any pixel DLSS adds that is no present in the native image will be counted as an error, even if the overall picture is better than the native image.

In fact someone told him this, and it is the biggest flaw of such a comparison:
"There is also the issue that the native image can sometimes look worse. So you reference image should probably be rendered using supersampling anti-aliasing. Which you than can compare to native, FSR, DLSS."

 

[muon]

Can't believe what i read.
Especially in this game it is said that DLSS is making the image "better than native". That will mean by definition that any comparison with the native image will show errors in the DLSS screenshot. Heck FSR can be 99.99% as accurate as the native image and DLSS only 80% as accurate and DLSS image can still look better if the native image is bad.
So what is the purpose of such comparisons? Take the images with that bridge for example, assuming they are real. Any pixel DLSS adds that is no present in the native image will be counted as an error, even if the overall picture is better than the native image.

I did get that point. However, given the numbers were showing DLSS to be far better to begin with it wasn't that relevant. If anything a supersampled comparison would favour DLSS as you've explained.

What we really need to see is how nvidia does it when they train their models. They have to somehow create an identical 1440P input image for an equivalent 16K supersampled image. They also need to so this across time. I guess if you have complete access to the rendering it becomes easier.

 

[zx128k]

Hell have fun, mathlab code.

clear all; close all;clc
imdata = imread(Image 1 path for ref) <---- Right click copy as path and paste in windows 11.
figure
imshow(imdata) <--- Open window and show the image
title('Image Native') <--- Name of Image

imdata2 = imread(Image 2 path) <---- Right click copy as path and paste in windows 11.
figure
imshow(imdata2) <--- Open window and show the image
title('Image DLSS or FSR') <--- Name of Image

MSE = immse(imdata, imdata2) <--- work out MSE value
SSIM = ssim(imdata, imdata2) <--- work out SSIM value

The console window will tell you the result. Use 100% the same camera view in your sample, in the game of your choice. Goes without saying really, if you want a valid result. Remember this just compares image 1 and image 2. If the image is better than native this is a negative impact on the result.

 

[ICDP]

To try and get this thread back on track I though I would post a link to this.

FSR OpenVR
This basically uses the FSR upscaling techniques to improve performance in OpenVR games. I have tried it in a few VR games and it does exactly what is sets out to do and in many cases it is close to, or indistinguishable from native (other than the UI). Though there are some things to be considered first.

• Unlike FSR it does affect menus in the game. FSR by design should be added earlier in the pipeline before any UI items are added. This mod affects the whole screen so should not be considered how FSR would look if integrated into each game specifically.
• Some games will not work (Half Life ALyx being one)

https://github.com/fholger/openvr_fsr

I tried in Fallout4 VR, Skyrim VR, DCS and IL-2 and it works very well in these games giving my RTX 3080 a 25% - 30% boost with FSR Ultra Quality settings.

 

[Besty]

This thread reminds me of all the other NV vs. AMD OCUK 'post war' threads.

- Frame-pacing
- Adaptive Sync
- Mantle
- FXAA vs. MLAA

Every single time whilst Nvidia has the initial upper-hand, AMD gets it right eventually, they can't compete with the marketing muscle nvidia have but they do act in the right way for the whole of the community, given time.

NV also have plenty of shills available to keep posting on forums like these, I would give it 12 months and then call the victor. It won't be Nvidia simply on the basis that pretty much every new console game running will implement FSR.

 

[phatboy]

i have tried several games with FSR on the 6700XT now and it does make a pretty impressive difference. Godfall gains 30 odd FPS on the quality preset and still looks great.

Had a stab at the rift breaker demo yesterday (actually a pretty cool little game as a side note) and with RT on and everything to ultra it again makes a really nice jump with FSR switched on.