You stated it is broken, you got to prove it. Really is that simple. If you cant be bothered doing the mathlab stuff etc. Showing his program is wrong. Making sure the images are the same frames. How can I judge you. I never bothered.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
Fidelity Super Resolution in 2021
- Thread starter Grim5
- Start date
You stated it is broken, you got to prove it. Really is that simple. If you cant be bothered doing the mathlab stuff etc. Showing his program is wrong. Making sure the images are the same frames. How can I judge you. I never bothered.
I can see you're not bothering.
After all the times I've mentioned the author discrediting their comparison themselves you're saying its my call.
I can see you're not bothering.
After all the times I've mentioned the author discrediting their comparison themselves you're saying its my call.
Well it took me less than a minute in a program I have never used, so 10 minutes is lots of time. MSE is the correct method via wiki but SSIM is also calculated.
Test program, completed in Mathlab.
Two random images on my desktop. Lots of errors.
Same image zero, error.
MSE Error amounts. Same images as the forum post.
FSR vs native = 226.6046
DLSS vs native = 66.8661
Zero means the ref image and A are the same.
Source images.
Native
FSR
DLSS
SSIM Results
FSR SSIM = 0.8758
DLSS SSIM = 0.9530
The value 1 indicates the highest quality and occurs when A and ref are equivalent.
I confirm the authors conclusion that the DLSS sample is closer to the native sample image using his samples. In fact the DLSS sample is very close to the native sample. Mathlab results kind of shows DLSS is really good and should be taken as native quality based on these samples.
I have not checked the samples for issues, just used them to work out MSE and SSIM in Mathlab.
Last edited:
You are nitpicking.
They aren't discrediting what they wrote. There are ways to improve what they have done which is what they are fairly accepting.
Nothing anyone does is perfect. But that doesn't make their conclusion or evidence wrong.
The problem is that what has been written is beyond your comprehension so all you see if that final sentence about how they agree with more time it could be improved.
They aren't discrediting what they wrote. There are ways to improve what they have done which is what they are fairly accepting.
Nothing anyone does is perfect. But that doesn't make their conclusion or evidence wrong.
The problem is that what has been written is beyond your comprehension so all you see if that final sentence about how they agree with more time it could be improved.
Complete nonsense.
The initial numbers and claim were being posted by zx128k while somehow failing to note that it was rapidly followed by criticism that the author accepted.
Just like this and another 3 times before that.
https://www.overclockers.co.uk/forums/posts/34968578/
And now I'm being asked to fix the inaccuracy that makes his evidence useless.
What inaccuracy?
The reddit poster hasn't done anything inaccurately.
They've used MSE to compare a native image against a DLSS and FSR version of the image. Perfectly acceptable analysis.
If you want to use SSIM to disprove that analysis, then go ahead. I think you'll be disappointed as it will almost definitely arrive at the same conclusion.
The inaccuracy in creating data which zx128k was then posting as fact without its criticism which happened to be directly underneath but ignored.
It's not a criticism. You reading it as such is the silly bit here. The people posting below have said nothing to say the conclusion of the reddit poster is wrong.
They have suggested ways of strengthening the conclusion by using a more suitable error metric and a supersampled image. Do you even understand what they are saying and why?
edit:
I was being harsh given the analysis did turn out to be flawed.
Last edited:
Do you have any idea what you're arguing over.
That image is a highlight of a post from earlier where I demonstrated that the author was no longer happy with their post which zx128k happened to be posting around here.
https://www.overclockers.co.uk/forums/posts/34968608/
https://old.reddit.com/r/Amd/comments/om4910/marvels_avengers_fsr_vs_nvidia_dlss_comparison/
Amazing how everything isn't in a partial piece of information you looked at.
Soldato
Using MSE to compare jpgs, jpgs of screenshots of a game that arent all produced from exactly the same location/direction in game as well. That's flawed. really flawed.
But that isn't what has happened, or at least no one has shown that to be the case.
You have one guy pointing out even a slight difference in camera angle will be flawed. Which is perfectly acceptable if he then showed that to be the case.
You have also posted something about how a piece of analysis someone has done could be flawed without actually showing it was. If that was the threshold for discrediting anything, then we couldn't believe anything.
University A has done an experiment to show medicine X works well against illness Y. You can't just come up with how they might have messed it up with zero evidence. It's easy to throw mud without actually doing any work.
I've seen you post it ten times. You still don't get it and keep making up the lie the author is no longer happy with what he has done. The author has accepted there are ways to improve his analysis.
Last edited:
Please see my above post, I have reached the authors conclusions, using mathlab. DLSS has a SSIM of 0.9530 when compared to native and FSR has a SSIM of 0.8758. Also FSR has a MSE of 226,6046 and DLSS a MSE of 66.8661. He is correct that his samples show DLSS is almost completely the same image as the native image.
Lol, you've put in a lot of effort to try win an internet argument. At least you aren't just throwing uninformed words though.
Soldato
That's exactly what happened, you only have to read the thread on reddit. The 4k native, FSR and DLSS screenshots were all taken from slightly different perspectives, the FSR screenshot being the most different. Then the author took the same section of each image (which were jpegs when they should really be uncompressed) and used MSE to compare them. It's totally flawed.
You know how that python script works, right? It pixel peeps. If you are drawing a scene from 3 slightly different perspectives then the results will reflect that perspective skew which is no good if you are trying to measure the difference in render quality. It needs to be dead on exactly identical perspectives or it wont work. and comparing jpegs also means the results include any compression artefacts.
Its 4 lines of code really.
Okay fine. I'll actually take a look if you have.
edit:
I've seen the larger thread with people talking about it now.
The compression artefacts should cancel out so not worried about that.
However, the images are indeed at very slightly different camera angles which makes it flawed. Looking at the images, even if they were dead on the same angle, the scene is almost definitely not completely deterministic with random differences throughout which also makes it flawed, although you could argue even that would cancel out.
Good idea though.
Last edited:
So when the author accepted the criticism that the post they made (that you were reposting) was lacking...
...you took the exact same same authors samples and repeated what the author did and got the same conclusion and you don't see a problem with all that being the same as the author did.
How did that fix any of the problems that were brought up about the methodology in the first place?
What a joker, yesmanning for zx128k but not once looking at the material til now.
You can compare compressed images. You just need a reference and an image to compare with it. The three image are basically the same or the results would be off. The DLSS result is far too close to the native to be too different. The FSR image is just not as good but close. I would like to have control over the source images but they are all close to the native image.
Soldato
So you haven't bothered? Why are you arguing then?
Jobastion comments on Marvel's Avengers FSR vs NVIDIA DLSS comparison (reddit.com)
No you cant. No they aren't. And yes..they are off. Use the imagediff tool and take a look at the 4k FSR screenshot vs DLSS and native and you'll see how the perspective is off. It invalidates the test as does the use the jpg images. You cannot pretend the use of lossy compression doesnt affect the final output.
The author was not wrong, you were. The authors program gets different values but the conclusion is the same. DLSS sample is almost completely the same as the native sample.
So you haven't bothered? Why are you arguing then?
Jobastion comments on Marvel's Avengers FSR vs NVIDIA DLSS comparison (reddit.com)
No you cant. No they aren't. And yes..they are off.
You can thats what MSE is for, see the wiki article I linked too. Both MSE and SSIM values are provided.