madVR - high quality video renderer (GPU assisted)

[Dion]

regsvr32 just hangs on Win10 (in VirtualBox).

Seeing this as well. But I am just running windows 10 normally.

[sneaker_ger]

For those of you with keyboard shortcuts and NGU not working, please try this test build:

http://madshi.net/madVR91b.rar

Does it fix the problem?
Yes. (both shortcuts and NGU luma doubling working now)

[Siso]

For those of you with keyboard shortcuts and NGU not working, please try this test build:

http://madshi.net/madVR91b.rar

Does it fix the problem?

Working here :)

[madshi]

Glad to hear that. So how's performance and quality compared to NNEDI3 for you?

[Siso]

Glad to hear that. So how's performance and quality compared to NNEDI3 for you?

On my poor GTX 550 Ti, the rendering times are 55 ms...in short pretty heavy. Super-xbr 75 or 100 runs very smooth tho.

[TheShadowRunner]

Fixed here as well :thanks:

[noee]

For those of you with keyboard shortcuts and NGU not working, please try this test build:

http://madshi.net/madVR91b.rar

Does it fix the problem?

Yes madshi, fixed here. Win7 Ult

[aufkrawall]

looks like NGU doesn't like IVTC images and MPEG2 compression artifacts:
doubling into spline3 AR

NGU medium: https://img3.picload.org/image/rdwlpoao/ngu.png
XBR100: https://img3.picload.org/image/rdwlpoap/xbr100.png
source: https://img3.picload.org/image/rdwlpoac/source.png
I'm having a really terrible quality cartoon sample whichs shows something similar:
http://www42.zippyshare.com/v/5ggfF2qv/file.html
(careful, it's 2 frames, not one)
It's a really good test to show some weaknesses of a chroma scaler regarding ringing and aliasing, btw (NGU medium is doing very well in this regard and high just ever so slightly better).

[imhh11]

we just have to copy these 2 files in the madvr folder ??
if yes, it didnt work for me.

Win10 64 bits V1607

[madshi]

regsvr32 just hangs on Win10 (in VirtualBox).
Yes, I'm aware of the problem. For now, you can just overwrite the files of an older madVR build. There's generally no need to "install". Just overwriting the files usually works just fine.

Performance:
R9 GPU
chroma and image doubling NGU High: 20s
chroma and image doubling NNedi 64: 22s
This is what I like to see. Now the question is why huhn has so much worse results on his AMD 480.

looks like NGU doesn't like IVTC images and MPEG2 compression artifacts:
doubling into spline3 AR

Hmmmm... That's ugly. And NNEDI3+SuperRes4 has no problems with that?

I noticed that AMD video color settings are not applied in fullscreen mode when 1080p video is playing on my 1080p screen. The only difference is that there is a "Crop 2" procedure applied during the render.
If I use MPC HC zoom feature and outzoom the color settings are applied. Is there any way to make it work ? I use a combination of Madvr and AMD video settings, works fine with all videos except 1080p.
AMD video settings are usually not good quality. Why are you using them at all? I think they're only applied if you use either DXVA scaling or DXVA deinterlacing.

we just have to copy these 2 files in the madvr folder ??
if yes, it didnt work for me.
Those 2 files are supposed to fix a very specific problem with keyboard shortcuts not working. They're not supposed to fix the installer not working. To workaround that, overwrite 90.24 with all the 91.0 files and then overwrite them with these 2 files on top of that. Sorry for the inconvenience.

[imhh11]

Those 2 files are supposed to fix a very specific problem with keyboard shortcuts not working. They're not supposed to fix the installer not working. To workaround that, overwrite 90.24 with all the 91.0 files and then overwrite them with these 2 files on top of that. Sorry for the inconvenience.

shortcuts are still not working for me.
Windows 10 64bits 1607

edit: NGU doesnt seems to be working ... :(
Also why does it say 30hz ? my tv is set to 60hz..
https://s16.postimg.org/9773dpted/ngu.jpg https://s15.postimg.org/ik14ofkcb/evr.jpg

[huhn]

i'm just asking and hope it is not true. is there a possibility that NGU needs a ton of Vram?
GPU queue of 4 (i'm using 8 normally)

i need 3.5 gb Vram with NGU high and i need 2.5 gb Vram with XBR 100.

edit:
the MPC-HC screenshoot function took the wrong source frame for what ever reasons.
and it was NGU med -> XBR 100 -> spline 3 AR

i forgot that quadrupling was a thing...
and nnedi3 256 with SR 4 doesn't look better: https://img1.picload.org/image/rdwlwgdg/nnedi3sr4.png
they look like clones...

here the correct source screen: https://abload.de/img/source2beugl.png (i will replace it in the comparison post too )
i guess the problem comes from using the old render path. just a guess the new render path is still broken on win 10.

[AngelGraves13]

I'll stick to ReconSoft for Chroma, as I think it still looks the best, especially with noisy encodings. I'm looking forward to an updated NGU for Chroma with Luma assist.

Image Doubling with NGU, just like NNEDI3 is too taxing and pointless for 1080p->1440p. Jinc + SR3 upscaling looks well enough and keeps my render times under 25ms on my 1080 GTX. NGU-High shoots it up to 60ms, causing massive framedrops. It might be useful for 480i content.

[cyberscott]

@madshi

Below are some quick runs comparing NGU to NNEDI3 64.
NGU does look real good so far on 1080p videos. Will check lower resolution later.

Running Nvidia 1080 gtx and Windows 10 64 bit
1080p23 to 4kp23
Debanding Med/High
Ordered Dithering
image enhancements and upscalling refinements not used.
Image upscalling Lancroz4 AR

Image Doubling NGU High, Chroma upscalling NGU High, 39ms (get drop frames with this one, of course!)
Image Doubling NGU High, Chroma upscalling NGU Med, 30ms
Image Doubling NGU Med, Chroma upscalling NGU Med, 14ms

Image Doubling NGU High, Chroma upscalling Superxbr 29ms
Image Doubling NNEDI3 64, Chroma upscalling NGU Med, 20ms
Image Doubling NNEDI3 64, Chroma upscalling Superxbr 16.5ms

[nevcairiel]

looks like NGU doesn't like IVTC images and MPEG2 compression artifacts:
doubling into spline3 AR


Many "sharp" scalers have that problem, as they don't know whats image and whats artifact.

Would love some good mpeg denoiser and de-blocker for those old encodes sometimes to run before scaling, but i'm too lazy to manually activate them on a file-by-file basis =p

[huhn]

yeah you don't want to see this without anti ringing...

[fedpul]

Hi madshi. Thanks for the new build. Talking about IQ and performance for Chroma Upscaling S-XBR is the king, NGU is not good at Chroma Upscaling compared to Recon Soft or S-XBR. Now for doubling I must say WOW!!! It is incredibly good but incredibly taxing too, for some references, I can double Luma and Chroma a 1080p movie with NGU Medium, and downscale it using SSIM 1D with render times about 30ms, but with NGU High render times go past 100ms and I can't notice the difference in IQ. NGU High is most suitable for SD Content where it shines and the performance hit for doubling a DVD movie is between 20ms and 25ms. Hope this can help.

EDIT: I am using a GTX 970 @1400/1950.

[fedpul]

yeah you don't want to see this without anti ringing...

Totally agreed!!! Anti-ringing is hyper neccesary.

[aufkrawall]

NGU is not good at Chroma Upscaling compared to Recon Soft or S-XBR.
Sample? From my tests, the opposite is true.

[Warner306]

Installer not working for me either. Will wait for an updated build.

[fedpul]

Sample? From my tests, the opposite is true.
I test it with one frame from the movie Howl's Moving Castle (DVD Release). I can upload samples tomorrow.

Enviado desde mi MI 4LTE mediante Tapatalk

[aufkrawall]

Thx.

Two examples for chroma s-xbr vs. NGU:

NGU:
http://abload.de/thumb/ngud6al8.png (http://abload.de/image.php?img=ngud6al8.png)

s-xbr 100 AR:
http://abload.de/thumb/sxbrjxx5s.png (http://abload.de/image.php?img=sxbrjxx5s.png)

With this sample, s-xbr has problems with kind of "inner haloing" or "outer darkening" (how you like to put it).

NGU:
http://abload.de/thumb/ngu24cxik.png (http://abload.de/image.php?img=ngu24cxik.png)

s-xbr 100 AR:
http://abload.de/thumb/sxbr2l5an6.png (http://abload.de/image.php?img=sxbr2l5an6.png)

Here s-xbr is quite blurry and there is a lot of dark haloing.

[trip_let]

Maybe it's just me but I'm wondering if a lighter value of "soften edges" and also "add grain" could be offered (between off and 1). When scaling very low-resolution content the differences can be fairly visibly substantial. There may be some utility in a smaller nonzero setting.

Still testing but NGU looks great to me so far.

[aufkrawall]

I agree that finer steps for soften edges could be beneficial for some users.
I myself got used to the sharp edges quickly, though. :)
As long as there is no bloating...

[ryrynz]

Here s-xbr is quite blurry and there is a lot of dark haloing.

You know the whole thing is blurry right? Not a good test for chroma. Nearly every chroma upscaler has dark haloing and super-xbr is one of the least haloing upscalers along with NNEDI3. When NGU is updated to be guided by luma it probably will do it too. What you're seeing here I think appears beneficial because of the type of content and how poor the resolution is in the same way bilateral would. I'm expecting super-xbr to be eclipsed by NGU for chroma when the upgrade comes.

[jkauff]

Just a couple of quick observations on NGU, madshi:

I can easily double 480p and 720p sources at high quality with my GTX 1060, but on 1080p even the "medium" setting gives me about one frame drop per second. I've been using NNEDI3 32 neurons with no problems. So don't get rid of NNEDI3 until a slightly more efficient version of NGU is available.

Also a caution. NGU works great on D3D 9, but I get massive frame drops on everything at both NGU strengths when using D3D 11. This is on Win 10 Anniversary Edition. Not an issue for me, but if people run into the same problem, this is why. Maybe for now you should disable D3D 11 if NGU is in use.

Love the results, though! Great work.

EDIT: 1080p now works with no drops with NGU medium. I just had to turn off SuperRes (which was at 1).

[ryrynz]

Also a caution. NGU works great on D3D 9, but I get massive frame drops on everything at both NGU strengths when using D3D 11. This is on Win 10 Anniversary Edition. Not an issue for me, but if people run into the same problem, this is why. Maybe for now you should disable D3D 11 if NGU is in use.


What content? What are your stats? What else you running with NGU?

[jkauff]

What content? What are your stats? What else you running with NGU?
480p, 720p, 1080p. I changed nothing except turning off D3D11, and then everything works fine with NGU. Could be a bug in NGU, or a Win 10 problem. I doubt if madshi did much (if any) testing on Win 10 build 1607. No idea if the same problem exists on Win 8.1.

[ryrynz]

No idea if the same problem exists on Win 8.1.

I haven't played with it on my 960 yet but try a few earlier driver versions. I'll post my experience with it shortly.

[aufkrawall]

You know the whole thing is blurry right?
And why would it be better with even blurrier super-xbr?

[ryrynz]

And why would it be better with even blurrier super-xbr?

super-xbr is far from blurry or soft for chroma if using ~100+

[Asmodian]

NGU looks very nice! I really like NGU High doubling with animated content.

Nvidia Titan X (Pascal) 375.95, Windows 10 x64 14965.rs_prerelease.161104-1700, I also used the test build of madVR (91b). Final resolution 2560x1440.

Default settings except using windowed overlay, ED1 dithering, and a 3DLUT. Chroma upscaling = Bicubic75, Image upscaling = Lanczos3, Image downscaling = Bicubic100. This is total average rendering times without any empty files with funny names. ;)


Titan X (Pascal)

Luma doubling
1920x1080p24 (2038 MHz)
NNEDI3 256 - 28.6 ms
NNEDI3 128 - 16.7 ms
NGU High - 18.0 ms

1920x1080p24 (1430 MHz)
NNEDI3 64 - 13.8 ms
NNEDI3 32 - 9.9 ms
NGU Med - 9.7 ms

1280x720p24 (1430 MHz)
NNEDI3 32 - 5.8 ms
NNEDI3 64 - 7.5 ms
NNEDI3 128 - 11.4 ms
NGU High - 12.0 ms
NGU Med - 5.6 ms

Luma and Chroma doubling
1280x720p24 (2038 MHz)
NNEDI3 128+128 - 19.5 ms
NGU High+High - 21.6 ms

[aufkrawall]

super-xbr is far from blurry or soft for chroma if using ~100+
Compared to NGU, it of course is.

[ryrynz]

First thoughts on NGU.. Jesus it's sharp. Thins lines more than I like.. I find I have to use a soft scaler like softcubic 50 to tone it down to something that actually looks natural, but I'll play with the edge softening as well and see how I go.

Yeah soften edges is a must for lower resolution animated content I can stick with Jinc when using soften edges.

The question is if I can make things look better than my anime setup current combo, lot's of screenshots tonight I think..

Compared to NGU, it of course is.

See above comment.


EDIT: 1080p now works with no drops with NGU medium. I just had to turn off SuperRes (which was at 1).

Yeah I figured you had something else running..

I have no issues with D3D11 and NGU on either setting on my 960.


We'll I've reached a conclusion, I won't be replacing NNEDI3 with this at least for animated content. If an option comes around to it being a bit more relaxed in style like NNEDI3 then I'd take another look. The edge smoothing helps and so do some of the softer scalers when used in conjunction with it but it's not enough to improve how it creates lines, it just helps mask the issues.

Yes, NNEDI3 is softer but I can work with that through post processing it doesn't deform lines as much as NGU and is cleaner to boot which to me is far more important than anything else.


1) Do we still need NNEDI3 image doubling?


Yup.

[heiseikiseki]

For those of you with keyboard shortcuts and NGU not working, please try this test build:

http://madshi.net/madVR91b.rar

Does it fix the problem?

for my environment Ctlr+J still doesn't work.

Windows 10
Nvidia GTX860m(optimus)

[Anima123]

for my environment Ctlr+J still doesn't work.

Windows 10
Nvidia GTX860m(optimus)

Me neither, nVidia GTX880m, Windows 8.1.

[QBhd]

for my environment Ctlr+J still doesn't work.

Windows 10
Nvidia GTX860m(optimus)

There is something odd about the fixed file... when unpacking, the RAR wouldn't overwrite files until I clicked yes to remove the archive after unpacking. Very strange never witnessed this before. Maybe the same thing is happening to you? Ctrl+J does work fine on my Windows 10 machine once I figured this out.

QB

[Anima123]

Me neither, nVidia GTX880m, Windows 8.1.

Finally, I figured out in which circumstances the OSD will reject to show.

When run mpc-be directly, ctrl-J will show the OSD correctly. However, if I start mpc-be via 'Run with graphics processor' with my GTX880m, the OSD failed to show when ctrl-J pressed.

This is quite weird behavior I never met before. BTW, I do have Optimus with this laptop, if it means something to the developer for debugging.

[heiseikiseki]

There is something odd about the fixed file... when unpacking, the RAR wouldn't overwrite files until I clicked yes to remove the archive after unpacking. Very strange never witnessed this before. Maybe the same thing is happening to you? Ctrl+J does work fine on my Windows 10 machine once I figured this out.

QB
I unpacked it and then copy and past them to the madvr directory(overwrite).
But't CTRL+J still doesn't work.

[cnckane]

AMD video settings are usually not good quality. Why are you using them at all? I think they're only applied if you use either DXVA scaling or DXVA deinterlacing.


Thanks, I will pass that then.

Btw. I tried this new NGU algorithm. Does anybody else notice a high pitching noise coming from the VGA when this algorithm is enabled ? It's like when you are in a game and there is 500+ FPS in the menu.

[madshi]

shortcuts are still not working for me.
Do you have an Optimus system?

Do you have any 3rd party software running which might also access the GPU, like f.lux, FRAPS, GpuZ or something like that? Or maybe some sort of debugger or something?

i'm just asking and hope it is not true. is there a possibility that NGU needs a ton of Vram?
GPU queue of 4 (i'm using 8 normally)

i need 3.5 gb Vram with NGU high and i need 2.5 gb Vram with XBR 100.
With which source video size? NGU does consume some more Vram, that's true. But it should be a "fixed" cost, meaning it should be independent of the GPU queue size.

edit:
the MPC-HC screenshoot function took the wrong source frame for what ever reasons.
and it was NGU med -> XBR 100 -> spline 3 AR

i forgot that quadrupling was a thing...
and nnedi3 256 with SR 4 doesn't look better: https://img1.picload.org/image/rdwlwgdg/nnedi3sr4.png
they look like clones...

here the correct source screen: https://abload.de/img/source2beugl.png (i will replace it in the comparison post too )
i guess the problem comes from using the old render path. just a guess the new render path is still broken on win 10.
I'm not completely sure how this related to the other problems you mentioned? Does this have anything to do with the high VRam usage or the slow render times?

About slow render times: Are you using any custom flush settings? Have you tried resetting madVR to default settings as a test?

Below are some quick runs comparing NGU to NNEDI3 64.
NGU does look real good so far on 1080p videos. Will check lower resolution later.
Thanks!

NGU High runs slower for you than expected. But not as slow as huhn reported.

Now for doubling I must say WOW!!! It is incredibly good but incredibly taxing too, for some references, I can double Luma and Chroma a 1080p movie with NGU Medium, and downscale it using SSIM 1D with render times about 30ms, but with NGU High render times go past 100ms and I can't notice the difference in IQ.
The difference between NGU Medium and High in image quality isn't all that dramatic, so it's not a big problem if you can't use High. How does performance compare to NNEDI3?

Maybe it's just me but I'm wondering if a lighter value of "soften edges" and also "add grain" could be offered (between off and 1). When scaling very low-resolution content the differences can be fairly visibly substantial. There may be some utility in a smaller nonzero setting.
So just some lighter settings would do the trick? We don't need more settings in between the current settings, I hope?

I can easily double 480p and 720p sources at high quality with my GTX 1060, but on 1080p even the "medium" setting gives me about one frame drop per second. I've been using NNEDI3 32 neurons with no problems. So don't get rid of NNEDI3 until a slightly more efficient version of NGU is available.

Also a caution. NGU works great on D3D 9, but I get massive frame drops on everything at both NGU strengths when using D3D 11. This is on Win 10 Anniversary Edition. Not an issue for me, but if people run into the same problem, this is why. Maybe for now you should disable D3D 11 if NGU is in use.

Love the results, though! Great work.

EDIT: 1080p now works with no drops with NGU medium. I just had to turn off SuperRes (which was at 1).
SuperRes doesn't seem to be beneficial for NGU, anyway.

Are you saying that the option "use D3D11 for presentation" makes problems when using NGU, but makes no problems when using other image doublers?

NGU looks very nice! I really like NGU High doubling with animated content.

Nvidia Titan X (Pascal) 375.95, Windows 10 x64 14965.rs_prerelease.161104-1700, I also used the test build of madVR (91b). Final resolution 2560x1440.

Default settings except using windowed overlay, ED1 dithering, and a 3DLUT. Chroma upscaling = Bicubic75, Image upscaling = Lanczos3, Image downscaling = Bicubic100. This is total average rendering times without any empty files with funny names. ;)


Titan X (Pascal)

Luma doubling
1920x1080p24 (2038 MHz)
NNEDI3 256 - 28.6 ms
NNEDI3 128 - 16.7 ms
NGU High - 18.0 ms

1920x1080p24 (1430 MHz)
NNEDI3 64 - 13.8 ms
NNEDI3 32 - 9.9 ms
NGU Med - 9.7 ms

1280x720p24 (1430 MHz)
NNEDI3 32 - 5.8 ms
NNEDI3 64 - 7.5 ms
NNEDI3 128 - 11.4 ms
NGU High - 12.0 ms
NGU Med - 5.6 ms

Luma and Chroma doubling
1280x720p24 (2038 MHz)
NNEDI3 128+128 - 19.5 ms
NGU High+High - 21.6 ms
Thanks for the detailed speed comparison. Once again, NGU High performs slower than expected. It does perform near to NNEDI3 64 with my 7750, so I expected similar results with higher performing GPUs, too. Maybe I do have to use D3D11 for NGU (when it's available) to speed things up, I don't know...

First thoughts on NGU.. Jesus it's sharp. Thins lines more than I like.. I find I have to use a soft scaler like softcubic 50 to tone it down to something that actually looks natural, but I'll play with the edge softening as well and see how I go.

Yeah soften edges is a must for lower resolution animated content I can stick with Jinc when using soften edges.

The question is if I can make things look better than my anime setup current combo, lot's of screenshots tonight I think..
Weren't you the one who was using AviSynth to thin lines? And now you say NGU thins likes *too much*? Sounds surprising. Do you have some screenshots for me?

Generally, NGU was optimized for upscaling content that was downscaled by using gamma light from the original highres master. If the studio (or custom encode) actually downscaled things in linear light, NGU will produce lines that are too thin. So maybe that's what's happening here. I was planning to offer a "linear light" switch for NGU at some point, which will be optimized for images that were downscaled using linear light. But it'll be some time before I get to that.

In all my tests with downscaling a "groundtruth" image (in gamma light, though!) and then upscaling it with NGU, NGU always produced lines in exactly the right width. They were never too thin. Are you sure you aren't just used to bloated lines?

Yes, NNEDI3 is softer but I can work with that through post processing it doesn't deform lines as much as NGU and is cleaner to boot which to me is far more important than anything else.
Strange. Of course NNEDI3 alone without SuperRes is much much softer (and thus "cleaner") than NGU. But if you apply SuperRes level 4, which is necessary to get as near to the "groundtruth" as possible, NGU is in my tests pretty much always cleaner and more truthful than NNEDI3+SuperRes4.

Finally, I figured out in which circumstances the OSD will reject to show.

When run mpc-be directly, ctrl-J will show the OSD correctly. However, if I start mpc-be via 'Run with graphics processor' with my GTX880m, the OSD failed to show when ctrl-J pressed.

This is quite weird behavior I never met before. BTW, I do have Optimus with this laptop, if it means something to the developer for debugging.
Oh, that's very interesting! I'll create a special log build to analyze this.

[ryrynz]

Btw. I tried this new NGU algorithm. Does anybody else notice a high pitching noise coming from the VGA when this algorithm is enabled ? It's like when you are in a game and there is 500+ FPS in the menu.

That's called coil whine (http://www.overclock.net/t/1094728/possible-fix-for-coil-whine), it's related to the quality of the design/parts in your graphics card. Best to avoid whatever manufacturer made that card in future.. you could try overclocking it, it may help.

[Sunset1982]

Hey Madshi!

Thank you very much for the new madvr version.

I did some quick performance test:

Test System: i5 6600, 16GB RAM, RX480 8GB, Crimson 16.11.4, Windows 10, madvr 0.91.b

File: Avatar 1920x1080p, 23,976, 8 Bit, MKV file upscaled to 3840x2160p



Old Settings:
Chroma: SuperXBR 125, AR
Luma: image doubling Superxbr 25 AB
+ Super Res 2
= 28ms


Changed Luma Doubling to NNEDI64, no Super Res:
=30ms


Changed Luma Doubling to NGU High:
=84ms


Changed Luma Doubling to NGU Med:
=36ms


Setting Chroma and Luma doubling to NGU Med:
=60ms

Setting Luma and Chroma to both use doubling @ NGU Med:
=97ms




Using Chroma Rconstuction Soft and Luma NGU med:
=36ms
Using Chroma Rconstuction Soft and NNEDI64:
=32ms

Using Chroma Rconstuction Soft and NNEDI128:
=61 ms


Conclusion:

NGU is to performance hungry for my system. I can only use it with medium setting, with chroma set to SXBR or RecSoft

[madshi]

Ok, two RX480 users now with very bad performance results. I smell a pattern. So my current impression is:

1) NVidia: a bit slower than expected.
2) RX480: a *lot* slower than expected.
3) All older AMD generations: just as fast as expected.

Will do some tests on my NVidia soon. Don't have an RX480 to test with, though.

[ryrynz]

Weren't you the one who was using AviSynth to thin lines? And now you say NGU thins likes *too much*? Sounds surprising. Do you have some screenshots for me?


Thin them to a point, NNEDI3 tightens them to some degree but it's not enough to look high definition so I take it further but Awarpsharp achieves is consistent and natural and works very well with NNEDI3.
I'll post some up later.


Generally, NGU was optimized for upscaling content that was downscaled by using gamma light from the original highres master. If the studio (or custom encode) actually downscaled things in linear light, NGU will produce lines that are too thin. So maybe that's what's happening here. I was planning to offer a "linear light" switch for NGU at some point, which will be optimized for images that were downscaled using linear light. But it'll be some time before I get to that.


Best the screenshots do the telling, along certain lines things thin where they shouldn't. NNEDI3 keeps the lines thickness consistent and without NNEDI3 enabled you can see that that's how they should be but NGU just seems to like
really constricting particular areas so that they no longer match the original line thickness and I don't like that.


In all my tests with downscaling a "groundtruth" image (in gamma light, though!) and then upscaling it with NGU, NGU always produced lines in exactly the right width. They were never too thin. Are you sure you aren't just used to bloated lines?


Pretty sure, I like things as sharp as they can be without looking unnatural, fat lines aren't nice but over thinned lines are worse.


Strange. Of course NNEDI3 alone without SuperRes is much much softer (and thus "cleaner") than NGU. But if you apply SuperRes level 4, which is necessary to get as near to the "groundtruth" as possible, NGU is in my tests pretty much always cleaner and more truthful than NNEDI3+SuperRes4.


I've seen impressive results on the forum, that Mononoke image upscale is amazing, but I don't get the same results with my content and thus haven't found any reason to run SuperRes, it just ends up looking over sharpened even with the lowest setting.

[madshi]

Do you have a "groundtruth" to compare to in any of your tests? Maybe an old SD encoding of some anime for which a Blu-Ray is now available?

[huhn]

With which source video size? NGU does consume some more Vram, that's true. But it should be a "fixed" cost, meaning it should be independent of the GPU queue size.
1080p. it looks a lot better since the reset.
it is most likely independent of the GPU queue it was still very high

I'm not completely sure how this related to the other problems you mentioned? Does this have anything to do with the high VRam usage or the slow render times?
when i'm running out of Vram i will get high rendertimes too so maybe.
GPU Vram usages in GPU-z is not reliable it could be out of ram even with numbers like 3.6 GB of 4 GB.


About slow render times: Are you using any custom flush settings? Have you tried resetting madVR to default settings as a test?

resetting settings helped as far as i can see. the flush settings where default i tried different settings later.
ShowRenderStep works like a handbrake for NGU.

FHD -> UHD 23p luma only
i get about 26-30 ms with NGU med now not sure how much NGU is eating. NGU high is using ~100 ms

[madshi]

resetting settings helped as far as i can see. the flush settings where default i tried different settings later.
ShowRenderStep works like a handbrake for NGU.

FHD -> UHD 23p luma only
i get about 26-30 ms with NGU med now not sure how much NGU is eating. NGU high is using ~100 ms
Yes, ShowRenderStep does slow things down, which is why it's a "hidden" option and shouldn't be turned on by default.

You reported earlier 22ms for NGU med. So it seems to be even slower now?? Or are you testing a different resolution now? The most important thing is not the speed by itself, but the speed compared to NNEDI3, to have a good reference point. So how do NGU med/high compare to NNEDI3 now?

[aufkrawall]

Btw. I tried this new NGU algorithm. Does anybody else notice a high pitching noise coming from the VGA when this algorithm is enabled ? It's like when you are in a game and there is 500+ FPS in the menu.
Yes, I got this too. Normally madVR causes just some electric buzzing and not this high "coil whine" sound, which is very strange.

[huhn]

Yes, ShowRenderStep does slow things down, which is why it's a "hidden" option and shouldn't be turned on by default.

You reported earlier 22ms for NGU med. So it seems to be even slower now?? Or are you testing a different resolution now? The most important thing is not the speed by itself, but the speed compared to NNEDI3, to have a good reference point. So how do NGU med/high compare to NNEDI3 now?

that was NGU only without any other renderstep using showrendertimes because everyone has different chroma scaler and stuff like that. didn't worked out...

FHD -> UHD normal rendertimes
NGU med 25-30 ms
nnedi64 32 ms
NGU high ~80-100ms
nnedi3 256 110ms
nnedi3 128 55ms


NGU med is about 10 ms faster now since i disabled showrenerstep and reset all settings. NGU high doesn't seem to care.