Does anyone here use NVIDIA image sharpening, or other upscaling techniques?

[Mark Kaine]

Just tried it, going from 1080p to something like 1600x900... I can see how it's beneficial for low power systems to get at least a stable experience, but boy is it blurry... 

Tried with and without gpu scaling, there's no difference - though I'm not really sure what gpu scaling even does. 

 

So is anyone actually using this and how do you like it? 

 

And what alternative image sharpening / upscaling options are there? 

 

Is DLSS similar, if not what makes it better? 

 

I just can't imagine an upscaling algorithm that isn't blurry?

[skimmilk5]

Are you talking about image sharpening in games or for scientific research in artificial intelligence and machine learning? They're two completely different workflows.

 

I use Nvidia's CUDA and Tensor cores to sharpen still images to aid in my research, but I suspect this isn't what you're talking about...

[Mark Kaine]

4 minutes ago, kimsejin5 said:

Are you talking about image sharpening in games or for scientific research in artificial intelligence and machine learning? They're two completely different workflows.

 

I use Nvidia's CUDA and Tensor cores to sharpen still images to aid in my research, but I suspect this isn't what you're talking about...

Well I was talking about games obviously. 

 

I remember when that Nvidia image sharpening thing came out it got a lot of praise... And after trying it out myself now, I do wonder why... 

 

Like I said I get it for low power systems but overall this doesn't seem to do anything special, you can do the same by forcing full screen in windows, it's just more practical and automated with nvidia 'image sharpening' I guess. 

 

And with how everyone is praising DLSS currently I think it is probably an improved version of this, and I'm just asking about experience with either technique, or with alternatives?

 

I know you can make low resolution photos and videos appear more sharp, I guess you really need those tensor cores or similar for that, as conventional upscaling techniques always produce artifacts and despite names like "sharpening" and "checkerboard rendering" the results have been generally always poor and a native image is usually preferable as it's just more clear (despite the lower resolution technically) 

 

 

 

 

[Pitbull Tyson]

Im playing on a 4k screen at 2k so I get the blur.  So this image sharpening will make that better?

[skimmilk5]

Just now, Mark Kaine said:

Well I was talking about games obviously. 

That's obvious? It all depends on who you happen to run across, and you happened to run across me, and the first thing I thought was image post-processing using Tensor. Forgive me for that misunderstanding.

 

Nvidia DLSS (Deep Learning Super-Sampling) is certainly an exciting technology, and Nvidia is in a great position to run lots of training on their models, since a good portion of the machine learning platforms out there are running Nvidia GPUs and are built for CUDA and Nvidia. It seems kind of wasted (to me) to put the Tensor cores in the consumer GPUs, since the training for technologies like DLSS are run ahead of time and deployed as software updates as pre-trained models, but I'm sure Nvidia has their reasons.

 

That being said. with time, DLSS can practically only get better! If you're using a higher-resolution display with a lower rendering resolution, if you can handle the processing overhead, then it should perform wonderfully. 

[Chiyawa]

36 minutes ago, Mark Kaine said:

Tried with and without gpu scaling, there's no difference - though I'm not really sure what gpu scaling even does. 

GPU scaling is a term describe how the GPU project the rendered image to the screen if the rendered image resolution is lower (in some case, higher, but that's rarely the case) than the monitor's resolution, or in a different ratio (for example, displaying a 4:3 image to a 16:9 monitor). There are only 3 settings available:

 

Centre - which the GPU will project the image in its resolution, so projecting a 640x480 image on a 1920x1080 monitor, you will only see a tiny image at the centre of the screen with black bars at top, left, right and bottom.

 

Preserve aspect ratio (some times known as best fit, maintain aspect ratio) - Same as centre, but it scaled the image up until 2 of the borders are gone. For example, if the rendered image resolution is 640x480 and the monitor resolution is 1920x1080, the GPU will enlarge the image until the top and bottom bar disappear, So the image will appear larger. There's still black bars at left and right of the display, however.

 

Stretch - Fill up the rendered image to the entire screen. A 4:3 image will look like it's being flattened on a 16:9 monitor. There will not be any black bars (unless the image is pre rendered with black bars).

36 minutes ago, Mark Kaine said:

And what alternative image sharpening / upscaling options are there? 

Image sharpening or upscaling just rendered the image to a higher resolution and scale down to your display resolution. Usually, depends on your monitor resolution, the image can appear sharper or become blockiness. This will also consume more VRAM, and unless the game has a very high resolution texture image, you won't notice the different.

36 minutes ago, Mark Kaine said:

Is DLSS similar, if not what makes it better? 

DLSS is a technique that GPU use to render texture and adjust the quality automatically. Usually, it will focus most of its performance to render image that is much nearer to your view and paying less attention to the object further from the view so it will be more responsive. This only works on 3D application, more specifically, 3D games.

36 minutes ago, Mark Kaine said:

I just can't imagine an upscaling algorithm that isn't blurry?

That really depends on how well your GPU upscales the image, the software that supports it, and your display. The ideal upscaling ratio is 2:1. For example, if your monitor resolution is 1920x1080, then you should tell the GPU to upscale the image to 3840x2160. Going lower resolution from your monitor resolution will result in blurry image.

 

Hope these explanation helps you understand on how GPU upscaling works.

 

Regards,

Chiyawa

[Chiyawa]

And by the way, because image sharpening uses more VRAM, those with GPU with low VRAM (2GB or 4GB) will have performance penalty when playing games. However, if you're watching Blu-Ray or DVD movies, it should not have much impact of your GPU.

[Helpful Tech Witch]

39 minutes ago, Mark Kaine said:

Is DLSS similar, if not what makes it better? 

 

dlss uses ai (which its comming from nvidia, so it has to be good) to upscale the parts of the image that mater (whats in the middle, text, players etc. Non important isnt. its really good on dlss 2.0 (not from own expirense)

[Mark Kaine]

3 minutes ago, Chiyawa said:

Hope these explanation helps you understand on how GPU upscaling works.

It does, partly, and thanks for taking the time to explain it all, but I do have a feeling you kinda mixed up down sampling and up scaling at some points (tho that could just be my reading comprehension) 

 

6 minutes ago, Chiyawa said:

That really depends on how well your GPU upscales the image, the software that supports it, and your display. The ideal upscaling ratio is 2:1. For example, if your monitor resolution is 1920x1080, then you should tell the GPU to upscale the image to 3840x2160. Going lower resolution from your monitor resolution will result in blurry image.

But then I don't understand the purpose of Nvidia image sharpening, it specifically tells you to use a lower resolution, that it then *up scales* to your monitors resolution, and indeed the result is very blurry - I think no different than if you'd let windows handle this (alt+enter to force full screen) Nvidia image 'sharpening' might be even worse... 

 

 

Tho I should probably use 720p for best results on my 1080p monitor (2:1)? 

[Chiyawa]

1 minute ago, Mark Kaine said:

But then I don't understand the purpose of Nvidia image sharpening, it specifically tells you to use a lower resolution, that it then *up scales* to your monitors resolution, and indeed the result is very blurry - I think no different than if you'd let windows handle this (alt+enter to force full screen) Nvidia image 'sharpening' might be even worse... 

 

 

Tho I should probably use 720p for best results on my 1080p monitor (2:1)? 

You should probably set your display setting to 1080p.

 

As for GPU image sharpening, if it allows you to adjust the image sharpening resolution, set it to 3840x2160.

[Warin]

I agree with Chiyawa. Your panel will always look best at its native resolution.

[Mark Kaine]

4 minutes ago, HelpfulTechWizard said:

dlss uses ai (which its comming from nvidia, so it has to be good) to upscale the parts of the image that mater (whats in the middle, text, players etc. Non important isnt. its really good on dlss 2.0 (not from own expirense)

Right, let's just say I'm *very* skeptical about that... Not that it doesn't do what it says, just that the result isn't a blurry mess (somehow only on my monitor of course, not in various 'independent' tests of course lol) 

 

The good news is probably that I can just brute force better image quality @ 1080p with a new GPU instead of having to rely on upscaling tech.

 

Im still very curious as at least dlss does sound promising... 

[Mark Kaine]

Just now, Warin said:

I agree with Chiyawa. Your panel will always look best at its native resolution.

Yeah i get that, but the purpose of nvidia image sharpening is to improve performance, without sacrifying image quality too much... 

 

Which, after my tests, is laughably wrong it just looks very blurry - tho I admit I didn't even think of using 720p which would make sense to test it properly. 

[AlwaysFSX]

11 minutes ago, Mark Kaine said:

Tho I should probably use 720p for best results on my 1080p monitor (2:1)? 

720p isn't a half ratio to 1080p.
960x540 iirc is 1:2

[Mark Kaine]

7 minutes ago, Warin said:

I agree with Chiyawa. Your panel will always look best at its native resolution.

OH yeah, except if you use super sampling... The only downside to that is you need the gpu power to do so but man does it look good, no artifacts or anything, say goodbye to aliasing also! 

[LogicalDrm]

-> Moved to PC Gaming