madVR - high quality video renderer (GPU assisted)

[huhn]

there new upscaling needs special implementation.

like many more of these new feature there are most tech demo for now.

nvidia did a really good job in featuring features and getting the hype up even through they can'T be used.

it looks like there upscaler DLSS needs training for the game first before it can be properly used.

[ryrynz]

It's a possibility that the tensor cores could be used to speed up some operations.. But that will likely be quite some time in the future if madshi even looks at it. I wouldn't expect this before 1.0 which is still some years away based on previous release history. I'd say the current upscalers in MadVR are pretty locked in for the time being.

[xabregas]

Maybe your GPU is getting too hot and clock speeds are being throttled? I see no other reason for rendering times to be OK for 5 minutes and then go up.Just a personal opinion, but chroma upscaling doesn't need a high precision scaler. NNEDI? Are you using a really old build?

Im using an old build because all this new updates are almost 4k and HDR oriented. Im playing with FHD TV mate. I have tried new versions but the new NGU is terrible at low, not so bad at medium and impossible at high. So should i use NGU instead of NNEDI with my hd7770?? And ventilation isnt a problem. Its 18 degrees celsius right now in my house.
But maybe some 3D setting enabled on amd control center?

He said he was and obviously is because as we all know it's been removed for some time.

Also yeah, don't bother with NNEDI3 for chroma.


So what should i use for chroma? Jinc? I dont like it. I like NNEDI3 more. Any improvements on NGU??

[huhn]

try super XBr on chroma i never understood nnedi3 for chroma.

and i highly doubt there is a huge difference between NGU aa low or med on chroma.

[hubi73]

Some Youtube audio clips (with still image) have a very low refresh rate of 6fps. It seems that this causing a delay in madVR's OSD response. For example: If I change the volume and hitting the arrow key or scrolling the mouse wheel there will be a delay of over one second before the OSD reacts. This makes it difficult to adjust the volume to a certain value.
I tested this with MPC-BE and Pot-Player.
I tested this with VMR9 too and there were the response times normal.

Clips with low refresh rate:
https://www.youtube.com/watch?v=2Lw7Vy9dVBo
https://www.youtube.com/watch?v=90SPoNnWHj4
https://www.youtube.com/watch?v=vKs01thUGUA
https://www.youtube.com/watch?v=G2ohR0xwrf8

In contrary here is an audio clip with normal refresh rate: https://www.youtube.com/watch?v=NxvSVpIN7Ys

http://jpegshare.net/images/b4/19/b4193fbe10b3c282f6a93b28d5bee5cc.jpg (http://jpegshare.net/)

[SirSwede]

I come from a plasma (Samsung PS42Q92HD) and I am moving to an OLED (Panasonic TX-55FZ800E).

I have been using Game Mode for madVR on the plasma, because it produced the best picture. To me, it is like madVR "likes" it best when the TV does at little as possible. Is it like that for just my eyes, or is it like that for someone/everyone else as well?

Anyway, moving to the OLED: Do I want to be using Game Mode with madVR on the OLED, just as I did with the plasma?

Will the madVR picture (and scalers) more come to it's right in Game Mode, than w.o. it?

:thanks:

[xabregas]

try super XBr on chroma i never understood nnedi3 for chroma.

and i highly doubt there is a huge difference between NGU aa low or med on chroma.

Well, it looks better than nnedi3 16 neurons and its only 10ms in both rendering times. Still far away from the PQ of nnedi3 32 neurons.

As for NGU it gives me high rendering times on high (more than 41ms). Medium is about 22ms but its kinda similar to nnedi3 16 in PQ. From what i saw NGU high gives amazing PQ. Shame i cant make it work. I need a new gpu im afraid. But i also need a 4k TV.

[mytbyte]

my eyes, or is it like that for someone/everyone else as well?

Anyway, moving to the OLED: Do I want to be using Game Mode with madVR on the OLED, just as I did with the plasma?

Will the madVR picture (and scalers) more come to it's right in Game Mode, than w.o. it?

:thanks:

Don't know about Panasonic OLED, but my Panasonic UHD LED can't do 24Hz in Game or Direct mode, it requires some processing...(Panasonic) plasmas are different, 24Hz works judder-free in any mode AFAIK. I guess plasmas in general were different beasts as they could be driven at any refresh.

[Asmodian]

I come from a plasma (Samsung PS42Q92HD) and I am moving to an OLED (Panasonic TX-55FZ800E).

I have been using Game Mode for madVR on the plasma, because it produced the best picture. To me, it is like madVR "likes" it best when the TV does at little as possible. Is it like that for just my eyes, or is it like that for someone/everyone else as well?

I think this is true on any display. You need to disable any internal processing that converts to 4:2:2, which is what Game Mode is probably doing. On my LG OLED I don't use Game mode, I use PC mode but it is probably the same effect.

[ryrynz]

Well, it looks better than nnedi3 16 neurons and its only 10ms in both rendering times. Still far away from the PQ of nnedi3 32 neurons.

The way you describe these makes it sound like there are huge differences between the various settings when there isn't. I've looked at all of these and the differences between NGU AA low and medium and Nnedi 16 and 32 are minor at best. NGU AA is a better chroma upscaler performance to quality wise than NNEDI3 is. I liked it but it was never worth it to have it set beyond 32 neurons. I'll happily upload comparisons showing NGU AA medium doing an overall better job than NNEDI3. This is why it was removed after all and I was a big fan of it up until it was removed.

[xabregas]

The way you describe these makes it sound like there are huge differences between the various settings when there isn't. I've looked at all of these and the differences between NGU AA low and medium and Nnedi 16 and 32 are minor at best.

I disagree! I can see the differences mostly in high action scenes like fights or even war movies. But close fights mostly. Ofc landscapes and faces or screenshots like we see on screenshot websites several times doesnt show any differences. :cool:

[ryrynz]

Chroma upscaling isn't affecting the picture as much as you think it is. Action scenes are not the instances where you'll see differences. You should back look through this thread for the comparison posts and screenshot yourself. Sounds like placebo.

[HDR]

I disagree! I can see the differences mostly in high action scenes like fights or even war movies. But close fights mostly. Ofc landscapes and faces or screenshots like we see on screenshot websites several times doesnt show any differences. :cool:
You have no idea what you're talking about.

You'll only notice chroma difference in things like red on black texts/backgrounds, and even then the differences are minor.

[ryrynz]

the differences are minor.

Depends on the scene, I've seen some quite large brightness changes on some scenes that are mostly mostly red/pink even without SuperRes enabled. Super-xbr 150 I believe is the brightest of them all.

[Betroz]

What can be the cause when there are a lot of repeated frames? My render queues are full and stats are around ~25ms. I read here in the thread that the 1809 update to Windows 10 could be the problem, so I rolled back to 1803 (using an older backup image). Same problems. Could it be the newest Nvidia drivers?

[ryrynz]

Think you'll find it's related to the refresh rate and the content you're running. Show OSD plz.

[Betroz]

Think you'll find it's related to the refresh rate and the content your running. Show OSD plz.

That seams more likely yes. I do have a new TV btw, so I will look into that :)

[HDR]

Depends on the scene, I've seen some quite large brightness changes on some scenes that are mostly mostly red/pink even without SuperRes enabled. Super-xbr 150 I believe is the brightest of them all.
It cracks me up that so many people spend so much time debating/arguing about chroma.

Yes, there are differences between algorithms, but gpu usage is more wisely spent in other areas.

I use Lanczos3 and I don't feel like I'm missing anything. I've compared all the algorithms extensively. It's the least important of all the scaling settings IMO.

IIRC madshi uses Bicubic 60 (the horror).

[SirSwede]

It cracks me up that so many people spend so much time debating/arguing about chroma.

Yes, there are differences between algorithms, but gpu usage is more wisely spent in other areas.

I use Lanczos3 and I don't feel like I'm missing anything. I've compared all the algorithms extensively. It's the least important of all the scaling settings IMO.

IIRC madshi uses Bicubic 60 (the horror).

I am still on Bicubic 75 and SoftCubic 80. :p

Although, I use Sharpen Complex 2 (post processing) in MPC-HC to bring out details in the picture - it works well with SoftCubic 80.

[huhn]

so he waste time and effort on chroma only scale like reconstruct?

and even if he uses bicubic 60 why should i care and do the same?

[SirSwede]

I think this is true on any display. You need to disable any internal processing that converts to 4:2:2, which is what Game Mode is probably doing. On my LG OLED I don't use Game mode, I use PC mode but it is probably the same effect.

So what you are basically saying - as I don't want to misunderstand you - is that when I buy an OLED: If it doesn't have PC Mode, then I want to be using Game Mode?

On the Panasonic TX-55FZ800E, they have arranged it so nicely, that Game Mode can be used on any picture mode. :D

[Betroz]

Show OSD plz.

Here is my OSD. It seams in order now, but i still get repeated frames when I pause og restart the video, but I guess that is normal.

https://imgur.com/a/rQYyGEO

[XMonarchY]

Any alpha/test builds of madVR Photoshop plugins planned for 2018 or even 2019? They would be SOO awesome! For example, current Photoshop tools that can de-noise well always leave ringing, which in itself, is pretty much noise. MadVR has excellent anti-noise WITH anti-ringing capabilities! I bet madshi could straight up sell madVR Photoshop plugins to Adobe for some $100K+!

[JarrettH]

The way you describe these makes it sound like there are huge differences between the various settings when there isn't. I've looked at all of these and the differences between NGU AA low and medium and Nnedi 16 and 32 are minor at best. NGU AA is a better chroma upscaler performance to quality wise than NNEDI3 is. I liked it but it was never worth it to have it set beyond 32 neurons. I'll happily upload comparisons showing NGU AA medium doing an overall better job than NNEDI3. This is why it was removed after all and I was a big fan of it up until it was removed.

I'd be curious to see NGU AA (on chroma) low vs medium. I'd like to reduce my settings from medium to low for less fan noise :D

[Alexkral]

The way you describe these makes it sound like there are huge differences between the various settings when there isn't. I've looked at all of these and the differences between NGU AA low and medium and Nnedi 16 and 32 are minor at best. NGU AA is a better chroma upscaler performance to quality wise than NNEDI3 is. I liked it but it was never worth it to have it set beyond 32 neurons. I'll happily upload comparisons showing NGU AA medium doing an overall better job than NNEDI3. This is why it was removed after all and I was a big fan of it up until it was removed.

This is why it was removed for chroma upscaling but not completely. I think there is a good reason to prefer it for image upscaling in some situations as I explained some posts before.
For chroma upscaling I use Super XBR and I don't think I need anything better.

[ryrynz]

Here is my OSD. It seams in order now, but i still get repeated frames when I pause og restart the video, but I guess that is normal.


Yeah that's looking good, you could reduce the queues unless for some reason you need them that high.

This is why it was removed for chroma upscaling but not completely.

Partially because it used OpenCL and madshi didn't really want it in madVR any more, but NGU AA was created to replace NNEDI3 and although it's not as sharp it does handle edges better.

I'd be curious to see NGU AA (on chroma) low vs medium. I'd like to reduce my settings from medium to low for less fan noise :D

You wouldn't notice the difference in motion, what you will find is that the color graduation will be smoother due to more sample points? and I find there is less ringing with higher settings.
Take a look along the edges of the folds of the cloak, especially the color graduation of bottom middle fold.

NGU AA low vs Medium (http://screenshotcomparison.com/comparison/123459) Bear in mind this is enlarged 300%.

The jump from NGU AA low to medium is greater than medium to high, personally I prefer medium as my base quality setting but it's relatively minor at the end of the day. I say change it to low and be happy with the lower noise.

[huhn]

there is a shader version of nnedi3.

[xabregas]

Depends on the scene, I've seen some quite large brightness changes on some scenes that are mostly mostly red/pink even without SuperRes enabled. Super-xbr 150 I believe is the brightest of them all.

Yeap, and if you enable adaptativesharpness by any reason it will be even brightest. Super Xbr is good but the brightness is a problem for me as green is not green anymore. everything green looks like sativa or lettuce. Back to nnedi3 again :scared:

It cracks me up that so many people spend so much time debating/arguing about chroma.

Yes, there are differences between algorithms, but gpu usage is more wisely spent in other areas.

I use Lanczos3 and I don't feel like I'm missing anything. I've compared all the algorithms extensively. It's the least important of all the scaling settings IMO.

IIRC madshi uses Bicubic 60 (the horror).

Im tempted in returning to Lanczo or Spline with adaptativesharpen enabled. But im not sure.

[huhn]

not what an chroma scaler has to do with adaptive sharpening and you should really read what ryrynz said.

and a chroma scaler will never turn something white is literal impossible.

[Betroz]

you could reduce the queues unless for some reason you need them that high.

I had my queue sizes at 8, but I got a lot of dropped frames. That went away when I put them higher. Why that is, I don't know.

[mytbyte]

I had my queue sizes at 8, but I got a lot of dropped frames. That went away when I put them higher. Why that is, I don't know.

On the contrary, in my setup, lowering queue size to 4 allowed me to use D3D11 native on a GT 1030 without stuttering or dropped frames...it has probably something to do with the gfx card having only 2GB of memory and more UHD frames in advance couldn't fit in there...

[ryrynz]

Might want to try setting the queues back to defaults again, that issue you had may have since been resolved. Only takes a moment to test.

[mytbyte]

Might want to try setting the queues back to defaults again, that issue you had may have since been resolved. Only takes a moment to test.

Sorry, failed to mention it's the case with HFR HDR (60 fps) in any scenario/settings. 24 fps, even when set to no downscale to simulate 4K output, smooth motion on and compromise off, it's smooth, albeit marginally (41 ms) with queue value of 16.

[Betroz]

Might want to try setting the queues back to defaults again, that issue you had may have since been resolved. Only takes a moment to test.

Was that a reply to me? I tested this yesterday with the newest version of MadVr and MPC-HC, so no nothing has been fixed.

[Betroz]

On the contrary, in my setup, lowering queue size to 4 allowed me to use D3D11 native on a GT 1030 without stuttering or dropped frames...it has probably something to do with the gfx card having only 2GB of memory and more UHD frames in advance couldn't fit in there...

Ok, but my card has 8GB of VRAM (GTX 1080).

[Ampallang]

Hi all,

can somebody help me I read plenty of sites of this great thread but unfortunately I am not that deep into all the madvr settings. I used the asmodian profile but I think there is plenty of optimization I can do. I run kodi with ds player and own a gtx 960 4g with a 4.5ghz devil canyon i5 windows 10 and 8gb ram. I am looking forward to Buy a new gpu soon but now where new cards will be released I think maybe it is a better solution to get the optimum out of my system and buy later cheaper an old rx580 or a vege 56/64.

My sources are dvd, blu-ray and uhd mkv I played on a calibrated LG oled. So does someone can provide me a good profile and settings for the gtx 960. I am so far happy with the results but when I read about the card should be able to run this and those my Asus strix won't be able to do this. But I think that will be related to my bad knowledge on topic. So if someone is willing to share his gtx 960 settings with me then I can take maybe a deeper look to understand why something won't work. As you can see I just used the lowest profile from user asmodian but when I read in this thread I have the feeling that I could achieve more and better but I have a wrong profile or so. I am on German region so 24/25/50/60p will be mainly the content to play.

Brgds
Ampa'

[el Filou]

and a chroma scaler will never turn something white is literal impossible.Not sure if it's brightness or saturation but I remember from my own comparisons between a 4:4:4 video and the 4:2:0 version processed by madVR, that for example ReconSharp made every red bit stand out more.
Classic scalers that don't take luma into account can't do that of course.

[huhn]

even reconstruct doesn't touch luma.

and to get a white pixel you need that in the first place and the chroma scaler has to remove the color.

and most important xbr 150 not using 150 but 100 instead is the real trick.

[Warner306]

My sources are dvd, blu-ray and uhd mkv I played on a calibrated LG oled. So does someone can provide me a good profile and settings for the gtx 960. I am so far happy with the results but when I read about the card should be able to run this and those my Asus strix won't be able to do this. But I think that will be related to my bad knowledge on topic. So if someone is willing to share his gtx 960 settings with me then I can take maybe a deeper look to understand why something won't work. As you can see I just used the lowest profile from user asmodian but when I read in this thread I have the feeling that I could achieve more and better but I have a wrong profile or so. I am on German region so 24/25/50/60p will be mainly the content to play.

Brgds
Ampa'

Try these:

4K UHD:

Chroma: NGU Anti-Alias (medium)
Downscaling: SSIM 1D 100% + LL + AR
Image upscaling: Jinc + AR
Image doubling: Off
Upscaling refinement: Off
Artifact removal - Debanding: Off
Artifact removal - Deringing: Off
Artifact removal - Deblocking: Off
Artifact removal - Denoising: Off
Image enhancements: Off
Dithering: Error Diffusion 2


1080p:

Chroma: NGU Anti-Alias (low)
Downscaling: SSIM 1D 100% + LL + AR
Image upscaling: Off
Image doubling: NGU Sharp
<-- Luma doubling: high
<-- Luma quadrupling: let madVR decide (direct quadruple - NGU Sharp (high))
<-- Chroma: let madVR decide (Bicubic60 + AR)
<-- Doubling: let madVR decide (scaling factor 1.2x (or bigger))
<-- Quadrupling: let madVR decide (scaling factor 2.4x (or bigger))
<-- Upscaling algo: let madVR decide (Bicubic60 + AR)
<-- Downscaling algo: let madVR decide (Bicubic150 + LL + AR)
Upscaling refinement: Off
Artifact removal - Debanding: medium/medium
Artifact removal - Deringing: Off
Artifact removal - Deblocking: Off
Artifact removal - Denoising: Off
Image enhancements: Off
Dithering: Error Diffusion 2


DVD:

Chroma: NGU Anti-Alias (low)
Downscaling: SSIM 1D 100% + LL + AR
Image upscaling: Off
Image doubling: NGU Anti-Alias
<-- Luma doubling: high
<-- Luma quadrupling: let madVR decide (direct quadruple - NGU Anti-Alias (high))
<-- Chroma: let madVR decide (Bicubic60 + AR)
<-- Doubling: let madVR decide (scaling factor 1.2x (or bigger))
<-- Quadrupling: let madVR decide (scaling factor 2.4x (or bigger))
<-- Upscaling algo: let madVR decide (Bicubic60 + AR)
<-- Downscaling algo: let madVR decide (Bicubic150 + LL + AR)
Upscaling refinement: Off
Artifact removal - Debanding: medium/medium
Artifact removal - Deringing: Off
Artifact removal - Deblocking: Off
Artifact removal - Denoising: Off
Image enhancements: Off
Dithering: Error Diffusion 2


You may not agree with the use debanding with those profiles. If you want the image to be sharper, try adding some image enhancements/upscaling refinement.

There is more information in the link in my signature if you are interested. The choice of what to use is mostly up to you.

[ryrynz]

albeit marginally (41 ms) with queue value of 16.

Cutting it terribly close there, you really want it under 40ms to avoid frame drops for 24fps content.

[SirSwede]

Try these:

4K UHD:

Chroma: NGU Anti-Alias (medium)
Downscaling: SSIM 1D 100% + LL + AR
Image upscaling: Jinc + AR
Image doubling: Off
Upscaling refinement: Off
Artifact removal - Debanding: Off
Artifact removal - Deringing: Off
Artifact removal - Deblocking: Off
Artifact removal - Denoising: Off
Image enhancements: Off
Dithering: Error Diffusion 2


1080p:

Chroma: NGU Anti-Alias (low)
Downscaling: SSIM 1D 100% + LL + AR
Image upscaling: Off
Image doubling: NGU Sharp
<-- Luma doubling: high
<-- Luma quadrupling: let madVR decide (direct quadruple - NGU Sharp (high))
<-- Chroma: let madVR decide (Bicubic60 + AR)
<-- Doubling: let madVR decide (scaling factor 1.2x (or bigger))
<-- Quadrupling: let madVR decide (scaling factor 2.4x (or bigger))
<-- Upscaling algo: let madVR decide (Bicubic60 + AR)
<-- Downscaling algo: let madVR decide (Bicubic150 + LL + AR)
Upscaling refinement: Off
Artifact removal - Debanding: medium/medium
Artifact removal - Deringing: Off
Artifact removal - Deblocking: Off
Artifact removal - Denoising: Off
Image enhancements: Off
Dithering: Error Diffusion 2


DVD:

Chroma: NGU Anti-Alias (low)
Downscaling: SSIM 1D 100% + LL + AR
Image upscaling: Off
Image doubling: NGU Anti-Alias
<-- Luma doubling: high
<-- Luma quadrupling: let madVR decide (direct quadruple - NGU Anti-Alias (high))
<-- Chroma: let madVR decide (Bicubic60 + AR)
<-- Doubling: let madVR decide (scaling factor 1.2x (or bigger))
<-- Quadrupling: let madVR decide (scaling factor 2.4x (or bigger))
<-- Upscaling algo: let madVR decide (Bicubic60 + AR)
<-- Downscaling algo: let madVR decide (Bicubic150 + LL + AR)
Upscaling refinement: Off
Artifact removal - Debanding: medium/medium
Artifact removal - Deringing: Off
Artifact removal - Deblocking: Off
Artifact removal - Denoising: Off
Image enhancements: Off
Dithering: Error Diffusion 2


You may not agree with the use debanding with those profiles. If you want the image to be sharper, try adding some image enhancements/upscaling refinement.

There is more information in the link in my signature if you are interested. The choice of what to use is mostly up to you.

Wow! Great recommendation!

Would you or anyone else, be able to do the same for me for 540p and 720p material on a calibrated Panasonic TX-55FZ800E 4K OLED, with a GTX 1050 Ti 4GB OC Edition Zotac card and I'd be eternally grateful! :p

(I don't like my image that sharp, since I will be adding Sharpen Complex 2 in MPC-HC (post processing) anyway.

:thanks:

[Alexkral]

there is a shader version of nnedi3.

You mean Shiandow's? It's NEDI not nnedi3 and anyway it doesn't seem very useful.

[LigH]

I already wondered ... Neural Networks in a GPU implementation?! Want! :eek:

[nevcairiel]

I already wondered ... Neural Networks in a GPU implementation?! Want! :eek:

Wut? Most Neural Networks rely on GPUs to run them at any speed, because GPUs are actually pretty good at that, because it can be paralized really well. Not to mention even newer stuff with specialized hardware just for that task, ie. Tensor Cores.

We had NNEDI3 in madVR for a while, a Neural Network based scaler. But it was removed because it was ultimately decided that the added complexity in madVR was not worth it (since it was the only component to use OpenCL IIRC), and other algorithms could flat out replace it at higher speeds and quality.

Neural Networks can be really powerful, and who knows if some day one might return to madVR, but the real challenge with Neural Networks is training them. You need absolutely massive compute power to train a Neural Network. All those fancy NVIDIA demos you see for image processing, those are trained on super computers, large clusters with hundreds if not thousands of NVIDIA GPUs to train the network. Of course no mere mortal has that sort of resources to train a network. :)

[nussman]

I already wondered ... Neural Networks in a GPU implementation?! Want! :eek:
What are you talking about? Of course nnedi3 by madVR was done by GPU ...

[videoh]

I already wondered ... Neural Networks in a GPU implementation?! Want! :eek: https://developer.nvidia.com/cudnn

[huhn]

https://forum.doom9.org/showpost.php?p=1736761&postcount=3202

yes in a shader. it wasn't slower to if i remember correctly.

[ryrynz]

Madshi will have a new release out soon with some big improvements to HDR tone-mapping.

[Warner306]

Wow! Great recommendation!

Would you or anyone else, be able to do the same for me for 540p and 720p material on a calibrated Panasonic TX-55FZ800E 4K OLED, with a GTX 1050 Ti 4GB OC Edition Zotac card and I'd be eternally grateful! :p

(I don't like my image that sharp, since I will be adding Sharpen Complex 2 in MPC-HC (post processing) anyway.

:thanks:

You could just replace every NGU Sharp entry with NGU Anti-Alias.

[Warner306]

Wut? Most Neural Networks rely on GPUs to run them at any speed, because GPUs are actually pretty good at that, because it can be paralized really well. Not to mention even newer stuff with specialized hardware just for that task, ie. Tensor Cores.

We had NNEDI3 in madVR for a while, a Neural Network based scaler. But it was removed because it was ultimately decided that the added complexity in madVR was not worth it (since it was the only component to use OpenCL IIRC), and other algorithms could flat out replace it at higher speeds and quality.

Neural Networks can be really powerful, and who knows if some day one might return to madVR, but the real challenge with Neural Networks is training them. You need absolutely massive compute power to train a Neural Network. All those fancy NVIDIA demos you see for image processing, those are trained on super computers, large clusters with hundreds if not thousands of NVIDIA GPUs to train the network. Of course no mere mortal has that sort of resources to train a network. :)

It was mentioned a while ago that NGU is based on neural networks. madshi posted at some point that the training was done by comparing downscaled images to the original. Maybe he has an evil lair with a bunch of super computers/GPUs to do this type of training?

If you look at screenshots, you can see that NGU picks up details like eyelashes that are missed by Lanczos and super-xbr. NGU very high pushes the GPU so hard that hundreds of thousands of repeated calculations must be taking place. NGU Anti-Alias is also a very similar upscaler to NNEDI3 and most seem to prefer it.

Personally, I think all images should eventually be scaled by neural networks when Tensor Core technology is scaled down. If, like me, you are frequently comparing Kodi VideoPlayer with Lanczos3 to NGU Sharp very high, they aren't the same image. Lanczos creates some artifacts, but it also smears a lot of valid detail. All of those eyelashes add up when the image is large.

With that said, I don't think NGU Sharp was trained well with poor material. I have seen some content look like an oil painting that kind of made me drunk while watching it.