madVR - high quality video renderer (GPU assisted)

[huhn]

why not NGU FAST instead of NGU low?

[Shiandow]

Probably because "NGU SLOW" doesn't have the same ring to it as "NGU high".

[huhn]

NGU complex
deal?

[madshi]

RX 480
FHD - UHD 23p
NGU low 19 ms
NGU mid 29 ms
NGU high ~85ms
super xbr 100 12 ms
nnedi3 32 22 ms
nnedi3 64 32 ms
nnedi3 128 46 ms
Still much slower than expected. NGU low was supposed to be almost twice as fast as NGU mid. It's clearly faster than NNEDI 16 on my PC.

Yes I am. Do you find it subtler or it only changes patter each frame now?
Were you using the highest setting before? v90.0 had 1-3, I think. v90.1 has 1-4, and the new 4 is a bit weaker than the old 3. Furthermore, the grain position is now different in each frame.

why not NGU FAST instead of NGU low?
I'm not sure about the "final" names yet. I can still try to scale NGU even lower/faster, or make steps between the current 3 options. Doesn't make sense to worry too much about the names before I've made up my mind how many quality/speed levels there are going to be.

[AndreaMG]

Were you using the highest setting before? v90.0 had 1-3, I think. v90.1 has 1-4, and the new 4 is a bit weaker than the old 3. Furthermore, the grain position is now different in each frame.
Before it was evident even at 1, with latest build the processing is not triggered no matter what the settings are in my case, I wonder why... FYI I use bilateral chroma scaling, NGU luma medium, JincAR, SSIM2D 100% :thanks:

[imhh11]

all working here!! thank you Madshi!
gtx1070
https://s14.postimg.org/g1h4j1pg1/all.jpg

[Smithy]

nice dynamic grain.

add grain 1 from v91.0 looks similar in v91.1 @ 3!

[aufkrawall]

With my 1070, NGU low is slightly faster than NNEDI3 16 (14.3ms vs. 14.8ms).

Yes I am. Do you find it subtler or it only changes patter each frame now?
More subtle, but with high strength still very noticeable.

[huhn]

I'm not sure about the "final" names yet. I can still try to scale NGU even lower/faster, or make steps between the current 3 options. Doesn't make sense to worry too much about the names before I've made up my mind how many quality/speed levels there are going to be.

i guess NGU *number* like super xbr than. but yeah doesn't matter yet.

[Smithy]

upscale JincAR SR2 vs. double NGU Med Image se1 ag3 / downscale SSim2D100
http://screenshotcomparison.com/comparison/191346

the add grain looks fantastic in this :)

madshi, can we use the add grain option @ Jinc upscaling or other scalers in next version or later ?


One thing that would be worth investigating:

Does SSIM downscaling produce any visible benefits over e.g. Bicubic150, when downscaling an image that was doubled before?

The reason I'm wondering is that one key benefit of SSIM is reproducing texture detail nicely. But when doubling an image, there's not much texture detail in the doubled image. So I think Bicubic150 might work just as well as SSIM in that situation. It's quite possible that SSIM only really helps when directly downscaling a high-res source. Anybody up for testing this in depth?


Bicubic150 AR vs. SSim2D100 AR
http://screenshotcomparison.com/comparison/191342

bicubic vs ssim looks similar to my eyes in this one, but the Screenshot isn't really Detailed ! ^^

[JarrettH]

How about Jinc AR + SuperRes (1-2) vs NGU low and med with catmull downscaling on real content?

I can't run NGU on my 550 Ti, but I can do the other. Curious as to what I'm missing. SuperRes with Jinc is more satisfactory to me than sxbr doubling.

[Thunderbolt8]

do I see this correctly that NGU is some kind of image doubling algorithm, but no normal upscaler? at least I cannot seem to choose it from the upscaling algorithms tab.

[JarrettH]

do I see this correctly that NGU is some kind of image doubling algorithm, but no normal upscaler? at least I cannot seem to choose it from the upscaling algorithms tab.

Yeah, I think the doubling implementation is becoming more important now with 1440p and 4k monitors

[theDongerr]

Hey new to the forum here and not sure where to start. Basically I'm running an R9 280 dual-x and pretty content with super-xbr 100 + SR 3 for upscaling 1080p24 -> 2160p24. I have been thinking about upgrading my GPU to take advantage of NNEDI3 or NGU. Will a GTX 1080 pretty much allow me to get the most of NNEDI3 and NGU for the purposes of doubling the 1080p image? I'm using uncompressed BD rips using MakeMKV.

I was originally looking at an asus gtx 1080 ROG strix card... seems like there are so many different flavors our there. Are there any that the board recommends?

[ispano]

NGU Feedback

Been testing NGU over the weekend on my RX 480. Rendertimes appear to be in line with what I expected and I'm sure it will see speed improvements along the way. As a metric, I can render NGU high >> luma doubling only >> 1080p to 4k @ under 41.7ms for film content. Driver is 16.11.4 clocked at 1425mhz and 2100mhz mem. NGU is a good middle ground for those who want the sharpness of xbr and the more "natural" look of NNEDI3. It's not a replacement for either IMO. There's also very noticeable aliasing that occurs when the dehalo algo is enabled in the artifact removal section. It's possible that NGU is already doing a lot of preprocessing so it's best to minimize as much filtering as possible.

1) Do we still need NNEDI3 image doubling? Or can NGU completely replace it?
Both XBR and NNEDI3 should never be an option for removal. They are "tried and true" so to speak from various other applications. NNEDI3 has been my reference since it's inception in avisynth. I like madvr because of its ability to provide a multitude of options. Each individual can find a configuration that works for them as well as their hardware.

2) Do we still need SuperRes (image doubling)? Or can NGU completely replace it?
SuperRes should stay. I think it works great for image upscalers like Jinc. Hardware requirements are far less with a good upscaler + SuperRes versus Image Doubling for resolutions like 1440p.

3) Do we still need NEDI image doubling?
If you absolutely had to clean house, then NEDI is the only candidate for removal. I personally prefer using it for 60fps with SR. It provides similar qualities to NNEDI3 but without the performance cost.

Finally thx for all your hard work Madshi and the great community feedback.

Side Note: Hardware Accelerated HDR playback works great on the RX 480. Must have been a driver issue with DXVA copy-back on initial release. Being able to use HDR made 3DLUT's from DisplayCal is a favorite addition of mine.

[Sp00kyFox]

tested a little bit more with NGU. very good job madshi, I'm pretty impressed with it. here is another comparsion:

http://screenshotcomparison.com/comparison/191345

[madshi]

can we use the add grain option @ Jinc upscaling or other scalers in next version or later ?
Maybe.

How about Jinc AR + SuperRes (1-2) vs NGU low and med with catmull downscaling on real content?

I can't run NGU on my 550 Ti, but I can do the other. Curious as to what I'm missing. SuperRes with Jinc is more satisfactory to me than sxbr doubling.
You're likely missing something. Make sure to disable chroma doubling and to pick catrom "image upscaling" to make "NGU-Low" fast. I'm not sure how often I have to say that. I think I've given this advice about 5 times on the last 2 pages... :(

Hey new to the forum here and not sure where to start. Basically I'm running an R9 280 dual-x and pretty content with super-xbr 100 + SR 3 for upscaling 1080p24 -> 2160p24. I have been thinking about upgrading my GPU to take advantage of NNEDI3 or NGU. Will a GTX 1080 pretty much allow me to get the most of NNEDI3 and NGU for the purposes of doubling the 1080p image? I'm using uncompressed BD rips using MakeMKV.
Your R9 280 might already be able to do "NGU-low" doubling with 1080p24 sources, just give it a try. Make sure you choose the other settings correctly. I'm not going to repeat them *another* time (see above).

[madshi]

Been testing NGU over the weekend on my RX 480. Rendertimes appear to be in line with what I expected and I'm sure it will see speed improvements along the way. As a metric, I can render NGU high >> luma doubling only >> 1080p to 4k @ under 41.7ms for film content.
Interesting! That seems to be much faster than what huhn is getting with his RX 480?

There's also very noticeable aliasing that occurs when the dehalo algo is enabled in the artifact removal section.
NGU already has a dehalo effect on its own, so I'd recommend to disable the separate dehalo option when using NGU.

SuperRes should stay. I think it works great for image upscalers like Jinc. Hardware requirements are far less with a good upscaler + SuperRes versus Image Doubling for resolutions like 1440p.
Have you tested "NGU-Low" with the recommended configuration for fast image doubling (see above, I posted this several times in the last 2 pages)? It's faster for me than Jinc+SuperRes!!

tested a little bit more with NGU. very good job madshi, I'm pretty impressed with it.
Glad you like it!

[JarrettH]

You're likely missing something. Make sure to disable chroma doubling and to pick catrom "image upscaling" to make "NGU-Low" fast. I'm not sure how often I have to say that. I think I've given this advice about 5 times on the last 2 pages... :(


No, sorry, I thought it was because I am using a 3D Lut and didn't mention that variable. Earlier I could swear the settings were linked and couldn't disable chroma.

Why does choosing catmull under image upscaling effect the speed? I thought that isn't used with doubling. I'm going 720p to 1080p

[MariaX9]

Here a couple of Numbers:

Chroma Upscaling: NGU High
Image downscaling: SIMM1D
Image doubling: luma and chroma NGU High
Image upscaling: Jinc
Render Time: 35s

Chroma Upscaling: NGU High
Image downscaling: SiSIMM1D
Image doubling: luma only NGU High
Image upscaling: Jinc
Render Time: 20s

Chroma Upscaling: NGU Medium
Image downscaling: SIMM1D
Image doubling: luma and chroma NGU Medium
Image upscaling: Jinc
Render Time: 17s

Chroma Upscaling: NGU Medium
Image downscaling: SIMM1D
Image doubling: luma only NGU Medium
Image upscaling: Jinc
Render Time: 12s

Chroma Upscaling: NGU Medium
Image downscaling: SIMM1D
Image doubling: luma and chroma NGU High
Image upscaling: Jinc
Render Time: 31s

Chroma Upscaling: NGU Medium
Image downscaling: SIMM1D
Image doubling: luma only NGU High
Image upscaling: Jinc
Render Time: 12s

Chroma Upscaling: NNEDI 64
Image downscaling: SIMM1D
Image doubling: luma only NNEDI 64
Image upscaling: Junc
Render Time: 22s

Chroma Upscaling: NNEDI 64
Image downscaling: SIMM1D
Image doubling: luma and chroma NNEDI 64
Image upscaling: Jinc
Render Time: 38s

Chroma Upscaling: super-XBR
Image downscaling: SIMM1D
Image doubling: luma and chroma super-XBR
Image upscaling: Jinc
Render Time: 6s

[fedpul]

Hi, I would like to share a comparison between some chroma upscalers: http://imgur.com/a/JSJBn

NGU looks too sharp and in some way it changes the brightness and other things, you can see how the pink stripes in the needle look thicker with NGU.

I will be doing render times comparison with 3 different movies. 480p24, 720p24 and 1080p24 with different doubling algos and bring it here in XLS format or plain text as you prefer.

Sorry for my late answer, I was very busy and did not have time to reply some things mashi and others told me.

[JarrettH]

Jinc with SR2 vs NGU low

http://screenshotcomparison.com/comparison/191350 (eh, not a great capture to use)

http://screenshotcomparison.com/comparison/191351

http://screenshotcomparison.com/comparison/191355

What do you think, madshi?

What I see are the same details with less overall coarseness which SuperRes tends to bring out, and contrasting edges draw less attention to themselves

Is NGU low a better doubler than super-xbr? I'm curious to know more :) What are the technical differences between low, med, and high? Is each setting the same sharpness?

[Warner306]

Just watching something now. Figured out how to use NGU medium quality at 720p -> 1080p.

Is NGU too sharp to anyone else? Maybe soften edges is a mandatory requirement?

[theDongerr]

Is there a preferred gtx 1080 card?

Sent from my Nexus 6P using Tapatalk

[Sunspark]

Hi, I would like to share a comparison between some chroma upscalers: http://imgur.com/a/JSJBn
NGU looks too sharp and in some way it changes the brightness and other things, you can see how the pink stripes in the needle look thicker with NGU.


I know what you mean.. I've been using bicubic 60 AR awhile for chroma on my intel hd 6000 gpu.. I find that primary colours, especially red, yellow too, are more saturated with it. Most noticable on anime, and video scenes that have pale reds present. Everyone here likes superxbr, but I find that it doesn't have the same richness of colour. The difference is subtle, you likely wouldn't notice it if one wasn't looking for it, except I was able to notice the difference on an relatively uncalibrated TN monitor. If you like, try that exact scene with bicubic 60 AR, I'd be curious to see how it looks.

[Sunspark]

@madshi

Hi, I am sorry I am late sending the log file to you. In the summer I had said that aspect ratio scaling seemed to select different values. Monitor's native resolution is 2048x1152 which is 16:9. Using a sample video it starts at 21:9 default aspect. Manually changing it to 235:100 changes it to 2.35:1 by extending the vertical slightly and circles looks more proportionate as a result compared to the initial setting. Some other files, e.g. tv, also have this aspect scaling subtly off, 16:9 v.s. 1.85 or some other number. In this file I started the video twice using both settings. https://dl.dropboxusercontent.com/u/50794070/Misc/madVR%20-%20log.zip

[Anima123]

I am glad to find that NGU low can be used to playback 576p -> 1080p with the old 640m. It is still slower than SuperXBR though, NGU low 22 ms while SuperXBR 16.6 ms.

The NGU low is more sharp and artifacts free than SuperXBR so a little bit more demanding is worth it.

Edit: Of course it would be great if NGU low can be further optimized to be in parallel the computing level of SuerXBR, if possible.

[Warner306]

Yes, I definitely prefer NGU with soften edges turned on. Cartoon is the word to describe it when left off. I didn't know too sharp was possible.

[Warner306]

@madshi

Is it possible to post a detailed description of NGU and what it does, similar to this post on NNEDI3:

Demonstration of NNEDI3 Image Doubling (http://forum.doom9.org/showpost.php?p=1663611&postcount=21592)

Of course, I'd like to link to it.

[Sunset1982]

Is there a preferred gtx 1080 card?


Good card come from MSI, ASUS or Palit. If you have enough room in your case I would go for MSI. They got a silent cooling solution.



@madhi:

Do you have any idea what causes the bad performance of NGU with RX480 cards? Driver issue or anytrhing other?


Any idea how to fix it? Is there anything I could do to help you find out?

[khanmein]

all working here!! thank you Madshi!
gtx1070
https://s14.postimg.org/g1h4j1pg1/all.jpg

y u didn't use D3D11?

[Ver Greeneyes]

I think I missed some important posts before, going to try turning off the deringing pre-processing next time and switch to a different option for image upscaling to minimize the overhead.

I wish there was a way to use a different regular upscaler for just the chroma doubling part though - if you have to double and then upscale more (e.g. 720p->1440p->2160p) I don't know if CatmullAR is the best option for the last part. I don't know how you'd work it into the UI though - the list for image doubling options is already pretty long. Maybe you could add at least one 'good' lower quality upscaler to the list of chroma doubling options?

By the way, I always feel weird neglecting Chroma like this, but I guess I really can't tell the difference anyway :P Oh, and I hope that question about whether SSIM after doubling makes sense or if bicubic150 is enough at that point gets answered - pretty easy to add a profile that just switches to bicubic150 if (targetWidth > srcWidth) || (targetHeight > srcHeight).

[Mistery 73]

good morning everyone ... with the new release it all works fine on win 10 madshi thanks ..
boys whom I recommended settings with the new NUG algorithms for upscaling from 1080-> 4K I have a video card 1070 GTX.

I want canditare Madshi the oscar award :)

[Anima123]

For me NGU is not just sharp, it is also clear in the edge of faces in video, even in motion.

SuperRes, IMHO, tend to add some kind of attachments to the edge of faces, which can be really painfully annoying. It's quite hard to sescribe the effect, maybe the algorithm was puzzled distinguishing the face edge and the background. I would've ask madVR to keep SuperRes if it didn't suffer that kind of artefacts.

[Gagorian]

I have an EVGA GTX 1070 on my HTPC, outputting to a 4k LG OLED.

I've replaced NNEDI64 (+SR2) for luma doubling of 720p/1080p to 2160p with NGU High and the performance is slightly better. There's a little bit of coil whine but it's inaudible with any audio on.

In case of 720p material what do you guys prefer, doubling 720p to 1440p with NGU High and using Jinc AR to scale to 2160p or quadrupling to 2880p with NGU High+Med and downscaling to 2160p using SSIM1D with LL and AR?

Currently I'm using sxbr 125 for all chroma upscaling. Does anyone else have another strong preference?

[ryrynz]

Currently I'm using sxbr 125 for all chroma upscaling. Does anyone else have another strong preference?

Recon soft and NNEDI3 32+ are the best alternatives. I prefer sxbr 125 AR.

[StinDaWg]

What are the technical differences between low, med, and high? Is each setting the same sharpness?
Interested in knowing this too. What are we actually giving up by using lower settings? Sharpness? More artifacts? Ringing?

I really don't see any difference between low-med-high when switching between them with normal live action video content.

High might look ever so slightly sharper than low on things like text, but that could just be placebo effect.

So, please tell us madshi so I don't lose my mind trying to spot the differences. :)

[HillieSan]

I have done another test with my RX 480.

Image upscaling: Calmull-Rom
Chroma doubling: off
Luma doubling:
NGU low 11 ms
NGU med 19 ms
NGU high 42 ms
NNEDI3 16 12 ms
NNEDI3 32 16 ms
NNEDI3 64 22 ms
NNEDI3 128 34 ms
NNEDI3 256 59 ms

Image upscaling to DXVA2 disables doubling and performs at 3-8 ms.

Enabling “run custom pixel shades in video levels instead of PC levels” saves 1-2 ms.

MPC-BE saves 1-2 ms compared to MPC-HC. (latest versions)

The strange thing is that I see just minor differences between DXVA2 and NGU low (and other scalers). The smoothness of DXVA2 (8 ms) makes the image better. Perhaps it is the noise in the movie that kills the improvements of the scalers. Or perhaps this is caused by the e-shift of the JVC X5000. Dunno.

BTW. The fan goes to high when using NGU low-high.

[Betroz]

Interested in knowing this too. What are we actually giving up by using lower settings? Sharpness? More artifacts? Ringing?

I really don't see any difference between low-med-high when switching between them with normal live action video content.

High might look ever so slightly sharper than low on things like text, but that could just be placebo effect.

So, please tell us madshi so I don't lose my mind trying to spot the differences. :)

I would like to know that too. NGU looks very similar to NNEDI3 64/128. Maybe that is a good thing, I don't know. One thing I have noticed though, is that there is more aliasing with NGU High than there is with NNEDI3 64/128.

So I will keep using NNEDI3 for now @ madshi

[aufkrawall]

Recon soft and NNEDI3 32+ are the best alternatives. I prefer sxbr 125 AR.
What's so great about reckon soft? Often there is no much difference to Jinc/super-xbr, unlike sharp it introduces dark haloing as well. So I don't see any point to use it over a non-luma guided scaler.
And still no one has provided images why super-xbr 125 would be the king for chroma. Did you try NGU low?

[fedpul]

What's so great about reckon soft? Often there is no much difference to Jinc/super-xbr, unlike sharp it introduces dark haloing as well. So I don't see any point to use it over a non-luma guided scaler.
And still no one has provided images why super-xbr 125 would be the king for chroma. Did you try NGU low?

Hi aufkrawall.

I did upload a comparison showing the differences a few posts ago.

This is the link: http://imgur.com/a/JSJBn

I compared: SXBR 100 AR, Recon Soft, NNEDI 256 and NGU High. From them all NGU is the worst to my eyes. NNEDI 256 is the best to my eyes but the performance cost is not acceptable for chroma, plus in motion is harder to see differences in chroma.

Where to look: http://imgur.com/SwKvVMg

[aufkrawall]

Thank you, I somehow missed that.
I suspect that we see the same problem here we see with really bad sources regarding NGU and luma doubling, the excessive ringing leads to quite unpleasant results.
In my American Dad cartoon example with not such excessive ringing, NGU clearly leads to better reconstructed chroma by even less ringing than s-XBR.
Can you try also NGU low? It's a bit softer than high.

[EpsilonX]

@madshi
Is this behaviour normal..?

http://i67.tinypic.com/992rfr.png

[aufkrawall]

Please uplodad the image to some imagehost, it won't be accessible here for some time else.

[fedpul]

Thank you, I somehow missed that.
I suspect that we see the same problem here we see with really bad sources regarding NGU and luma doubling, the excessive ringing leads to quite unpleasant results.
In my American Dad cartoon example with not such excessive ringing, NGU clearly leads to better reconstructed chroma by even less ringing than s-XBR.
Can you try also NGU low? It's a bit softer than high.

Hi, I will do it, I need to learn how to go to an exact frame. Cause I don't know how to do that, I will upload every algo again in the same frame. (Could be another frame but the same image) but it will be later. I am working in a big algo comparison. I will upload this comparison soon, I think it can help madshi.

[aufkrawall]

Can be a bit tricky to jump to the exact frames in mediaplayer, even when you tell MPC HC/BE to jump exactly to frame x. You may have to seek a bit frame by frame with ctrl + arrow keys.
A more "secure" way would probably be AviSynth with the Trim function.

[Gagorian]

Hi aufkrawall.

I did upload a comparison showing the differences a few posts ago.

This is the link: http://imgur.com/a/JSJBn

I compared: SXBR 100 AR, Recon Soft, NNEDI 256 and NGU High. From them all NGU is the worst to my eyes. NNEDI 256 is the best to my eyes but the performance cost is not acceptable for chroma, plus in motion is harder to see differences in chroma.

Where to look: http://imgur.com/SwKvVMg

I think sxbr 100 AR is very similar to NNEDI 256 in these examples. In fact so similar I doubt you could see the difference in motion.

[madshi]

I'm curious to know more :) What are the technical differences between low, med, and high? Is each setting the same sharpness?
It might be roughly the same sharpness, but the "higher" levels are cleaner, less aliased and more focused. There's no free lunch, after all.

Is it possible to post a detailed description of NGU and what it does, similar to this post on NNEDI3:

Demonstration of NNEDI3 Image Doubling (http://forum.doom9.org/showpost.php?p=1663611&postcount=21592)
You mean a detailed image comparison, explaining the strengths and weaknesses etc?

NGU looks very similar to NNEDI3 64/128.
No. NGU is *significantly* sharper than NNEDI3 64/128.

One thing I have noticed though, is that there is more aliasing with NGU High than there is with NNEDI3 64/128.
NGU in most cases has *less* aliasing than NNEDI3+SuperRes. A comparison to "straight NNEDI3" (without SuperRes) doesn't make too much sense because it's a completely different world of sharpness.

So I will keep using NNEDI3 for now
That's your choice, of course. But I think your comparison was probably flawed.

Did you use SuperRes in your comparison? If so, did you activate it for both NGU and NNEDI3? Do *not* do that. I've said it many times now: NGU does not benefit from SuperRes, results can actually get somewhat worse when switching SuperRes on with NGU. So use NGU alone, without SuperRes. On the other hand, if you want to compare NNEDI3 to NGU, you almost have no other choice than to use combine NNEDI3+SuperRes, because otherwise the difference in sharpness is simply too big to do any useful comparisons. So IMHO the proper comparison would be "NGU without SuperRes" and "NNEDI3 with SuperRes".

Is this behaviour normal..?
I don't know your settings, so I can't say for sure. I see nothing wrong with the screenshot, though.

[madshi]

RX480 users

I've received an email from a non-doom9 madVR user. He had big trouble with NGU performance with his AMD GPU, but found a way to fix it. Maybe it helps the RX480 users?

I think idle memory clock (=Radeon power management) is the root cause
of very slow NGU.

My Radeon 290X render times and memory clock are
16ms and 1250MHz when NNEDI3 64neurons
190ms and 150MHz when NGU high

and when I fixed memory clock to 1250MHz,
NGU render time goes down to 24ms. 8times faster!!

#PC setting
Windows10 Pro 64bit anniversary
AMD Radeon 290X & driver 16.10.3
MadVR 0.91.1 , 720p movie

[Neet009]

Something of my test.

GTX 1060 6GB (driver 373.06)
Monitor 1920x1080 60Hz
win7 x64
MPC-BE x64

The power state of videocard is troublesome, I have to set mpc-be64.exe to maximum performance on NV control panel, otherwise I would get variable rendering time(it would become very high) and many dropped frames.

madVR settings:
zoom control off
debanding on (low/high)
deringing off
smooth motion on
dithering - ED1
D3D11 FSE mode
chroma upscaling - Bilateral soft
image downscaling - SSIM 1D+AR+LL
image upscaling - Jinc+AR
image doubling - check double luma only except super-xbr

720x480 > 1920x1080
super-xbr 100 6ms
nnedi3 16 9.5ms
nnedi3 32 10ms
nnedi3 64 11ms
nnedi3 64+SR4 14.5ms
nnedi3 128 13ms
NGU low 9ms
NGU low 6ms(image upscaling - Catmull-Rom+AR)
NGU med 9.5ms
NGU high 11ms

640x480 > 1920x1080
super-xbr 100 6ms
nnedi3 16 8ms
nnedi3 32 9ms
nnedi3 64 10.5ms
nnedi3 64+SR4 12.5ms
nnedi3 128 14ms
NGU low 7ms
NGU low 5.5ms(image upscaling - Catmull-Rom+AR)
NGU med 8ms
NGU high 11ms

1280x720 > 2560x1440 > 1920x1080
super-xbr 100 8ms
nnedi3 16 11.5ms
nnedi3 32 13ms
nnedi3 64 17ms
nnedi3 64+SR4 24ms
nnedi3 128 24.5ms
NGU low 11ms
NGU low 9ms(image upscaling - Catmull-Rom+AR)
NGU med 13ms
NGU high 20ms

1920x1080 > 3840x2160 > 1920x1080 (image downscaling - bilinear)
super-xbr 100 10.5ms
nnedi3 16 19ms
nnedi3 32 23ms
nnedi3 64 31ms
nnedi3 64+SR4 52ms
nnedi3 128 50ms
NGU low 18ms
NGU low 12ms(image upscaling - Catmull-Rom+AR)
NGU med 22ms
NGU high 39ms

Acrodding to my test, for performance, NGU-high is between NNEDI3-64 and NNEDI3-128, NGU-med is as much as NNEDI3-32,
NGU-low is less than NNEDI3-16 but more than super-xbr 100.
In my opinion, NGU is amazing and useful, I think it is good for real-life content that I can raplace NNEDI3 and super-xbr, it brings the sharpness similar to NNEDI3+SR4 and less artifacts, and faster. But for anime content I still prefer NNEDI3(no SR) over NGU especially for low-res and high compression content. For anime, NNEDI3 did a good job of reducing aliasing, NGU is a little too sharp that aliasing and artifacts become more visible, NNEDI3 is more clean even though NGU might closer to original picture.(NGU-high has less artifacts than NGU-low, but still more than NNEDI3)
For anime, personally, I like NNEDI3 + thin edge1.0 + adaptive sharpen1.5(AB100%+AR).

A set of example:
originalhttp://i.imgur.com/D2y3ULU.png
NNEDI3 64http://i.imgur.com/obCswUn.jpg
NGU highhttp://i.imgur.com/GXZB2M7.jpg
NNEDI3 64+TE1.0+AS1.5 http://i.imgur.com/bRXAoKr.jpg