madVR - high quality video renderer (GPU assisted) [Archive] - Page 975

Asmodian

28th January 2018, 17:14

Are you watching 4K content on a display with a resolution higher than 4K? If so I would recommend NGU AA for chroma and NGU Sharp for luma. If your display is 4K or lower the options change.

Doubling is not useful unless you are upscaling, it does not automatically improve quality. There is nothing "stronger" than NGU and I would not say "super-xbr and NGU". NGU >> super-xbr.

d3rd3vil

28th January 2018, 17:16

NNEDI isnt even there anymore....ok so NGU. No I am watching the 4k stuff on a 4k display. Just want to maximize the quality. So NGU AA chroma and NGU Sharp for a 4k display doesnt do much? And there is the SuperRes filter and lots of stuff in upscaling refinement. But all that doesnt really do sth with 4k stuff on a 4k display?! Like it doesnt matter what to select under image upscaling like doubling or upscaling Jinc.....
So what exactly does improve the quality at least a bit? :)

el Filou

28th January 2018, 17:25

If you're watching 4K on a 4K display nothing is needed except chroma upscaling.
If you want to refine the image, look in processing => artifact removal/image enhancements.

d3rd3vil

28th January 2018, 17:29

If you're watching 4K on a 4K display nothing is needed except chroma upscaling.
If you want to refine the image, look in processing => artifact removal/image enhancements.

Yeah oki dok god bless you. Gotta look at the image enhancements settings mostly

mytbyte

28th January 2018, 20:37

Yeah oki dok god bless you. Gotta look at the image enhancements settings mostly

Just be mindful, madshi claims that if scaling is set to DXVA, chroma upscaling is also done in DXVA and setting there have no effect, if I understand correctly...

bcec

28th January 2018, 22:40

I find madVR's 'compress highlights' of HDR->SDR conversion to be too strong. It is trying to recover too much highlights, so much so that the result looks flat, without the necessary contrast.

For instance, a particular scene from Bladerunner 2049

bluray (http://i68.tinypic.com/14w4x11.png)
uhd with madvr HDR->SDR conversion (http://i66.tinypic.com/2vjue7d.png)
- 130nits
- preserve hue (hi, 50/50)
- compress highlights (measure peak lum, restore details, AB-100, AR)

It is worth noting that if I passthrough HDR to the display, what I see with my eyes is closer to bluray. To my eyes, highlights look bright and blown.

@madshi, would it be possible to tone down the strength of highlight recovery, or provide an option?

el Filou

29th January 2018, 00:16

Did you try 0% luminance reduction/100% saturation?

PS: please don't embed 1:1 images in the forum.

bcec

29th January 2018, 00:39

Did you try 0% luminance reduction/100% saturation?

0/100 is for hue, right? I don't see it making any difference for highlight compression.

PS: please don't embed 1:1 images in the forum.
Thanks. Changed to links.

mytbyte

29th January 2018, 00:48

Preserve hue makes RM_UHD 1000 nit 100% Amp 100% Sat calibration patterns very pale and of all kinds of different hues depending on choice, so the one who doesn't have an HDR TV doesn't really know which choice is closest to correct...

ryrynz

29th January 2018, 09:26

Upgraded from 960 to 1060, NGU high supersampling unlocked on 1080 content, nice.

steakhutzeee

29th January 2018, 14:29

Actually, the screen shot settings in madvr works with the default F5 key for screen shot of mpc be?

Axelpowa

29th January 2018, 20:54

Hi!

First of all I would like to thank you all the community that supports this software.

I have some doubts I'm not sure about.

First, which is better to use? Dxva2 native or copyback. I read that some options can only be used under copyback (image enhancements perhaps?)

Then, under ngu sharp there is an option called superresolution, but I don't see any difference in using at maximum strength or not using it.

My gpu is a geforce gtx 1080 msi armor, I5 7600 kabylake.

Regards!

Megalith

29th January 2018, 23:23

Can you link an image of that, what GPU card you're using? I'm using Intel iGPU Hd4600 to downscale 4k on 1080p display with DXVA and haven't noticed any jaggy.

I am using a 1080 Ti. Interestingly, DXVA looks considerably worse than even Nearest Neighbor:

DXVA:

https://i.imgur.com/OYXGyFE.png

Nearest Neighbor:

https://i.imgur.com/wNBBzbF.png

I wonder if there is something I should adjust in the NVIDIA control panel, or this is just how it is.

ryrynz

30th January 2018, 01:25

1080Ti and you want to use dxva? Why?

Asmodian

30th January 2018, 01:28

I wonder if there is something I should adjust in the NVIDIA control panel, or this is just how it is.

Are you using DXVA2 Native decoding in LAV?

austinminton

30th January 2018, 01:35

brazen1

30th January 2018, 02:07

You're not alone. Other quirky things too. MPC-BE, instead of quitting, replays and it's sporadic. Sometimes starting a title produces a black screen and many attempts later finally plays back perfectly. Other times, resetting 12 bit brings everything to life. Going to system image restore back to Jan 3rd prior to updating madVR, LAV, MPC -BE nightly, GPU driver, etc. Then slowly update one at a time and test a bit. If I get any results, I'll post back.

SamuriHL

30th January 2018, 02:37

I guess welcome back from vacation for me. Thanks, Microsoft. So I got a new insider slow ring build (17074) and now 3D playback is seriously broken for me. I've reverted the nVidia driver and J River MC 23 to previous versions that were working before I left for vacation and the problem is not resolved. Essentially, 3D playback plays fine but the screen goes completely blank when I try to bring up the player UI. I've tried in my normal windowed mode and FSE and both exhibit the behavior. If it's playing 3D, any time the player UI tries to draw it completely black screens me. Awesome. Just awesome. Anyway, I know we're not even going to think about trying to fix issues for in flight insider builds of windows but I just wanted to caution others that are in the slow ring that if you see this behavior, it appears to be something MS has done. Now to figure out what the hell I'm going to do to get myself out of this mess....do I forge ahead with a fast build or do I try to revert...decisions decisions. Maybe I'll throw darts at my tv and choose. LOL

el Filou

30th January 2018, 04:02

@austinminton I think madshi has said that the API doesn't provide him a way to set the display output bit depth.

ryrynz

30th January 2018, 08:31

I've not seen this option doing any harm and it saves a ton of performance.

Since we have real 4K content now, maybe I should activate the trade quality option "scale chroma separately" by default, because it will give a pretty large performance boost for 1080p displays, with only a very small quality loss?

Well I've looked at a few 4K videos and it does "harm" AFAIC, enough that I'd rather keep the option disabled and take almost 20 odd ms hit on my 1060. It's not like I'm doing anything with the 4K content other than downscaling anyway.. but with this option enabled as you can see in my previous screenshots it certainly affects the clarity of the image and based on what I've seen it's worth keeping it unticked if you have the performance to spare.
It's not just the general "chroma things" that get touched either like with upscaling..the whole image basically loses some vibrancy with this enabled.

invader

30th January 2018, 09:39

It seems to me the automatic bit depth switching in nvidia cp has stopped working as intended. Earlier if I would select 12 bit in 4k@23 RGB full and then switch to 4k@60, 4k@60 would work at 8bit and 4k@23 would continue to work at 12 bit. Now going back to 4k@23 (after a reboot) goes back to 8 bit. I like to keep my desktop at 4k@60. As an enhancement, would it be possible to allow to specify bit depth in auto display switching in madvr? That would remove all dependence on nvidia cp and its quirks.

Same problem for me. When using a custom 23p res, I seem to be able to switch to this res with 12bit after using a standard res with 12bit support first and then switch (e.g. 25p). Bit depth drop down in NVCP is actually blank at that point. But after some more switching to 60p@8bit the custom res resets to 8bit as well, really annoying. Only version that really worked is 385.28, which lets you actually switch bit depth in a custom res. That’s the way it should work imo.

Anyone know a solution to make the 12bit stick for a custom 23p res using the latest 390.65 Nvidia drivers?

mytbyte

30th January 2018, 10:05

I am using a 1080 Ti. Interestingly, DXVA looks considerably worse than even Nearest Neighbor:

I wonder if there is something I should adjust in the NVIDIA control panel, or this is just how it is.

I suppose DXVA scales the picture in 4:2:0 prior to any MadVR, and it prevents any MadVR chroma upsampling algorithms...

Sent from my GM 5 Plus d using Tapatalk

njfoses

30th January 2018, 12:41

Any feedback on new Nvidia driver 390.77?

oldpainlesskodi

30th January 2018, 14:00

Any feedback on new Nvidia driver 390.77?

Still has the same bug as described above by austinminton and invader, so for the moment I'm sticking with 382.53, which one of the later insiderbuilds shipped with.

K

arrgh

30th January 2018, 23:14

... it appears to be something MS has done. Now to figure out what the hell I'm going to do to get myself out of this mess....do I forge ahead with a fast build or do I try to revert...decisions decisions. Maybe I'll throw darts at my tv and choose. LOL

...or maby you go back to Win8.1 where you have virtually "full" control...
...and if the support of Intel GPUs would be as intensive as with NVidia or AMD (which are by the way also buggy as hell, when I see the driver-numbers exchanged for working with this but not with that...) then it would be heaven...
at least for machines which are meant mainly for media playback...

SamuriHL

30th January 2018, 23:53

...or maby you go back to Win8.1 where you have virtually "full" control...
...and if the support of Intel GPUs would be as intensive as with NVidia or AMD (which are by the way also buggy as hell, when I see the driver-numbers exchanged for working with this but not with that...) then it would be heaven...
at least for machines which are meant mainly for media playback...

Believe me I would if it were only for media playback. It has a couple other uses that need Windows 10, unfortunately, so I have to live with the consequences of that. In theory I completely agree with you.

LigH

31st January 2018, 00:35

I just experimented a bit with post-processing options ... and had to realize that activating the filter for compression artifacts (deblocking) is just too much for 1080p AVC on a Phenom-II X6 and 450 GT. But I saw only one core utilized. So I wonder if madVR is not even multithreaded, at least in a way that several modules would run in separate threads. Or is deblocking really such a heavy task? I believe ffdshow's implementation could handle that, even with additional random noise, but I don't have it installed anymore.

nussman

31st January 2018, 00:40

madVR uses GPU for video processing ...

LigH

31st January 2018, 00:47

JNW

31st January 2018, 00:59

On one laptop which is showing its age all I need to do is chroma upsampling. Mainly 720p. The highest this laptop can handle is spline64 (4 tap) w/ no AR. Is this what you guys would do or would you be concerned about aliasing and ringing. I see some people use lanczos but to me spline looks better on test patterns as it’s basically just as sharp but with less ringing. Even default bicubic60 I see more aliasing/stair-stepping than in spline but obviously spline has more ringing. I’m not sure there is a consensus on chroma, or is there?

ryrynz

31st January 2018, 01:10

Don't rely on test patterns. If it looks good to you then good enough. Try NGU AA low/medium.

LigH

31st January 2018, 01:14

Part of it is a matter of taste; I am not a great fan of "too sharp" upscaling, and the spline kernel is more moderate than Lanczos, especially with (less/fewer?) taps.

A pity the kernel visualization page of the ResampleHQ project (http://svn.int64.org/viewvc/int64/resamplehq/doc/kernels.html) disappeared.

I was so bold to upload a copy (http://www.ligh.de/resamplehq/kernels.html) for the preservation of useful information.

ryrynz

31st January 2018, 01:37

Yeah me neither, if I can tell an image has been sharpened then I'm generally of the opinion it's been oversharpened.
BTW your link is 404.

JNW

31st January 2018, 02:00

Very cool LigH. Thanks for that. I think spline 3-tap is the sweet spot for me. I like a sharper picture but obviously want to keep the artefacts down to a minimum. Strange because you don’t hear many people talk about spline when it’s a great algorithm in a lot of ways.

Ryrynz you have to use the second link. LigH was saying the first one is dead.

ryrynz

31st January 2018, 02:03

I'm aware, just a bit pointless linking it.
Yeah I use Spline 64 for anime through ffdshow before I awarp it, gives me a good result.

Asmodian

31st January 2018, 02:16

Spline 3-tap is an old favorite of many, it is simply too old to have been the sexy new thing for most so it doesn't get much publicity. :)

Edit: Thanks for hosting that page again LigH, I was also sad when it went offline. :thanks:

huhn

31st January 2018, 04:36

OK, you are right, I was too quick posting my question.

ProcessExplorer revealed that normal playback already takes my GPU to the limit, adding debanding doesn't matter much ... but deblocking goes beyond its capabilities, the playback stalls because the GPU gets saturated. Already using shaders for resampling is heavy, relying on DXVA2 mostly takes the GPU utilization back to just 55% (but deblocking is still utopical).

RCA is pretty much one of the hardest task madVR can do but unlike CPUs GPUs got so much faster that even a iGPU can easily outperform a old mid range GPU.

so utopical is not really the case.

but i was one of the old spline 3 AR user and i would still use it for "correction" scaling if it would be available not that it matters there.

and scaler like jinc, NGU AA and super XBR are not much sharper or sharper at all compared to spline they have a lot less scaling artifacts and have better AA abilities.
even NGU sharp is not an over sharpening scaler this can be proven by downscaling an image and upscaling it with NGU sharp and compare it to the source.

nevcairiel

31st January 2018, 10:32

Believe me I would if it were only for media playback. It has a couple other uses that need Windows 10, unfortunately, so I have to live with the consequences of that. In theory I completely agree with you.

Its your conscious chocie to use insider builds and forever suffer from it. No use needs insider builds. :)

SamuriHL

31st January 2018, 14:59

Its your conscious chocie to use insider builds and forever suffer from it. No use needs insider builds. :)

Of that I am aware. I wasn't asking anyone to fix it, just making others who may be going down the same path aware of the issues. :) It's still broken for me in the latest builds, as well. LOL But I am fully aware this is pain I brought upon myself by putting that machine in the insider program.

mclingo

31st January 2018, 20:15

@madshi - any chance you could revisit at refresh rate switching and 3D, the issue i'm always moaning about gets fixed if you disable refresh rate switching in MADVR and use KODI DS RFS instead, dont know if this is a bug or something you are aware off or whether its connected to the NVIDIA issues with 3D, worth a look though.

I much prefer MADVR switching as I also use MPC which has issues with RFS

-Hitman-

31st January 2018, 20:36

Hi, I've found a serious banding issue (scenes with white clouds show a pink posterization/banding effect) with UHD HDR movies, located the source to be DXVA2 - if I disable H/W acceleration (Native DXVA2) in the LAV video codec, then the banding is gone and the scenes look normal but start getting dropped frames w/out DXVA2.

I'm using an intel 8700k with on-board Intel 630 GPU, tried MB latest graphic drivers and also intel's latest offering but both exactly the same, anyone any idea's how to fix this issue?

Player MPC-BE
Latest MadVR
Latest LAV codecs - all used with UHD playback

Thanks.

Images...

https://image.ibb.co/dgzvDm/IMG_0659.jpg (https://ibb.co/cnyBm6)

https://image.ibb.co/nKvrm6/IMG_0660.jpg (https://ibb.co/j61mKR)

https://image.ibb.co/mwarm6/IMG_0661.jpg (https://ibb.co/cojTYm)

https://image.ibb.co/cwEj66/IMG_0662.jpg (https://ibb.co/nawxR6)

Asmodian

31st January 2018, 21:09

Do you get the same issue with DXVA2 copy-back or D3D11 decoding in LAV? DXVA2 native decoding is known to damage quality on Nvidia GPUs (maybe Intel has issues too?).

-Hitman-

31st January 2018, 21:18

Do you get the same issue with DXVA2 copy-back or D3D11 decoding in LAV? DXVA2 native decoding is known to damage quality on Nvidia GPUs (maybe Intel has issues too?).

I tested no H/W decoding, DXVA copyback and intel quick sync - all are clear but I get dropped frames with them all, Intel quick sync had the worst dropped frames, only DXVA native gave no DF but with the banding as mentioned..

sneaker_ger

31st January 2018, 21:20

Do you use DXVA scaling? (press ctrl+j)

-Hitman-

31st January 2018, 21:30

Do you use DXVA scaling? (press ctrl+j)

Chroma upscaling?

This is set in madvr to lanczos 3 taps +AR

Only downscaling I see (in madvr), is set for DXVA, this would not be used with 4k UHD on a 4K display.

Edit: Ctrl+j only mentions DVXA once - P010, 10bit, 4:2:0 (DXVA2)

https://image.ibb.co/e4rLeR/IMG_0669.jpg (https://ibb.co/c87tzR)

arrgh

31st January 2018, 21:53

@madshi
https://forum.doom9.org/showthread.php?p=1827545#post1827545
referring to my above post you answered
The decoder queue seems to be empty. Does it ever go higher than 1-1/16? If not, the decoder is your problem.

no, the queue never fills; so, nevcairiel came to the conclusion that an Intel driver bug prevents for those files the hardware decoding in LAV .83; so he issued .88 where the hardware decoding is disabled; but still problems remained : a choppy/jerky playback;

so, nevcairiel came to the conclusion that it is a renderer problem :
https://forum.doom9.org/showthread.php?p=1831863#post1831863
Software decoding is independent of the OS you use, or the drivers you have installed, so if you're seeing issues on one OS and not on another, its far more likely that this is caused by the rendering - which is outside of LAVs control.

only if I change the 3D setting in madVR from "auto" to "side-by-side" (as nevcairiel suggested) a smooth playback also of the faulty file-type is possible, but it’s very unstable: as soon I go with the mouse over the seek bar or fast-Forward, the playback gets jerky again;

so I have a few questions:

does the “side-by-side” trick provide the same quality as the “auto” mode (are 2 FHD pictures sent to the TV which it combines to a 3D-picture)?

How can it be that some files are played ok/better than others? All are basically h264 with ssif, right…?

I have noted a few differences between the working and not-working files here:
https://forum.doom9.org/showthread.php?p=1832220

(just as a hint: Some days ago, I have sent via PM a link to a sample of such a working/not-working sample)

sneaker_ger

31st January 2018, 22:05

Do you get the same issue with DXVA2 copy-back or D3D11 decoding in LAV? DXVA2 native decoding is known to damage quality on Nvidia GPUs (maybe Intel has issues too?).
I tested no H/W decoding, DXVA copyback and intel quick sync - all are clear but I get dropped frames with them all, Intel quick sync had the worst dropped frames, only DXVA native gave no DF but with the banding as mentioned..
LAV nightly (https://files.1f0.de/lavf/nightly/) adds D3D11 in case you haven't tested yet.

Asmodian

31st January 2018, 22:22

I tested no H/W decoding, DXVA copyback and intel quick sync - all are clear but I get dropped frames with them all, Intel quick sync had the worst dropped frames, only DXVA native gave no DF but with the banding as mentioned..

Yeah, DXVA2 native sucks. It is a little faster though. :(

Edit: sneaker_ger's point is important, D3D11 Native decoding is better than DXVA2 Native and also does not copyback, give it a try.