madVR - high quality video renderer (GPU assisted)

[madshi]

Originally Posted by ryrynz

Whenever I'm creating or deleting a profile the scroll bar resets back to the stop

What do you mean with "resets back to the stop"? I'm not sure.

Originally Posted by ryrynz

Would've liked for simple things like to to be fixed before upgrading to 0.91.7 but oh well

Well, fixing bugs often costs more time than improving and introducing features. If I had worked on all those bugs, there would be no NGU AA available right now.

Originally Posted by ryrynz

Now that I've had a quick play with NGU AA on the HTPC, I'm actually preferring it's more natural/balanced appearance at least in combination with my Avisynth profile during viewing.

Had to add some adaptive sharpen to balance things out but the overall result is quite nice, I'm now using very high for my anime 480 profile and medium for my 720.

So you're saying you're prefering NGU AA over NNEDI3 now? If so, I'd be happy to hear that.

Originally Posted by ryrynz

Madshi, would you consider releasing anything in between high and very high in future, there appears to be rather big jump between them in performance and appearance.

Yes, something in between high and very high is planned for the future, but it all costs serious development time, so I can only deliver things step by step.

Originally Posted by ryrynz

I'm also interested in looking at possibly any alternative builds of the very high setting, not sure if there are any tangents that can be looked at though.

Right now all the NGU algos are not perfectly fine tuned yet. I'm still doing rough adjustments right now. Once all the rough coordinates are dialed in, I'll fine tune all the algos and presets one by one. But that will take time, so don't expect it any time soon.

Originally Posted by Neo-XP

original (extracted by MPC-BE) :

super-xbr 100 / 125 / 150 :

NGU AA low / med / high / veryhigh :

Nnedi 16 / 32 / 64 / 128 / 256 :

super-xbr is not very good unfortunately compared to NGU AA and Nnedi.
Nnedi seems to be more precise (but not necessarily sharper) and hides artifacts better (for instance the artifacts under the black letters) than NGU AA. It also looks cleaner overall.

Again, I'm barely seeing a difference in these images. Maybe we need to collect better test images/samples for chroma upscaling. If I want to judge which algo is better, a test image should show something which jumps right into your eye.

Originally Posted by steakhutzeee

What about i want to learn how to properly configure Madvr last version? I'm on Intel i5 4590; an MSI R9 270X 2GB and 8GB ram.

Please try one of the many madVR guides.

Originally Posted by Oguignant

Hi. Can someone explain to me how "disable GPU gamma ramps" option works? I have 2 separate calibrations. One for madvr (3d lut file) and one for windows (matrix icc v4 file). If I enable the "disable GPU gamma ramps" option, does madvr ignore the windows calibration?

Yes, "disable GPU gamma ramps" should disable the ICC Windows calibration. It should be re-enabled when you close the media player. The best way to test something like this is to create crazy 3dlut and ICC calibrations which e.g. turn everything into a red haze or so. Then you can easily see what happens in which situations.

Originally Posted by Neo-XP

Is there any way to use NGU Sharp (med) for chroma doubling with NGU Sharp (high) for luma doubling ?
When I select "very high" for chroma doubling, it doesn't let me when "high" is selected for luma doubling.

I try to do this :

artifact removal : reduce banding artifacts medium & high
chroma upscaling : NGU Anti-Alias (high)
image downscaling : Bicubic150 + AR [relaxed]
luma doubling : NGU Sharp (high) 2x supersampling
chroma doubling : NGU Sharp (med) 2x supersampling
image upscaling : Jinc + AR
upscaling refinement : add grain 3
dithering : Ordered + colored noise + change dither

I don't think that NGU Sharp (med) for chroma doubling is a bad choice considering the rest of the configuration.
The limit for chroma doubling should be equal or at maximum one step below the luma doubling IMO, not two...

@madshi ? any chance to change this ?

Normally I would flat out say "no" because IMHO luma is *THAT* much more important that a 2 step difference sounds just right to me. However, NGU Med is only slightly slower than NGU Low, which makes me somewhat unsure about this.

Opinions, anyone else?

Originally Posted by 70MM

Ok I understand, I am upscaling 1080p >2160p so Upscaling refinement is actually working for me.

What I want to know then is, should I use Image enhancements or Upscaling refinement if I want to add a bit of extra sharpness?

Image enhancements applies sharpening before upscaling. Upscaling refinements afterwards. Generally, applying sharpening afterwards looks better but is less noticeable/effective, so you have to turn the sharpness dials up a lot more. Also applying sharpening after doubling is 4x as slow because it has to do the math on 4x as many pixels.

Generally, I wonder how much more sharpening you need when using NGU Sharp, thought? See also:

Originally Posted by Warner306

I'd use NGU Sharp, as it doesn't require added sharpening. If I wanted to add sharpening, I'd use SuperRes under upscaling refinement. But it isn't necessary with NGU Sharp.

^

Originally Posted by Ditto666

So for the first issue, I'm confused why changing any of the upscaling options as any effect AT ALL when the only possible configuration I could obviously have of the video is downscaling.

First of all, please don't make half your post bold or upper-case. That's like shouting, which is not nice.

See the madVR trade quality option "scale chroma separately, if it saves performance". You'll get slightly higher quality if you upscale chroma to 4:4:4 first, then downscale both luma+chroma at the same time. But if you want to save performance, you can use that trade quality option. Then madVR won't upscale chroma in your specific situation.

Originally Posted by Ditto666

For the second issue, how in the world would the anti-ringing filter when applied to anything (in this case only to the SSIM algorithm) result in the image being sharper? In this case, it's a significant difference too.

It shouldn't. If you can reproduce it, maybe you can post a comparison screenshot, with the OSD (Ctrl+J) being turned on?

Originally Posted by Ditto666

Lastly, never mind that it shouldn't do anything anyway given the mentioned setup, why does super sampling completely screw up and otherwise also have the opposite effect of blurring the image if using the NNEDI3 algorithm?

The purpose of supersampling is that you can apply some other algos in higher resolution. Usually you do supersampling, then you sharpen in the supersampled res, then you downscale again. Most of the sharpening algos look better when used this way. But of course it costs a lot of performance. If you just do supersampling without any additional sharpening, you just pile 2 resampling passes on top of each other, which is usually not beneficial, and may slightly reduce sharpness, depending on which exact algos you're using for supersampling doubling and downscaling.

Originally Posted by Ditto666

PS- I've had this question in mind for a while and was wondering if there was any real answer to it. With pretty much any scaling or other related options (excluding any image enhancements or sharpness refinement of course), do any of them actually add sharpness (as in contrast that's added between line edges) to the picture or in each case does it simply bring out more/less of the original image in its original resolution closer to if it were like a vector object? If it's the latter, would it be accurate to say that the sharper the image becomes, the better/more accurate it is, where it then just becomes a matter of lowering the sharpness on your display? I get that there can be parallel differences depending on the algorithm where increases in sharpness may be in different areas, but that aside... Obviously if I max everything out, the image appears to be overly sharp but, again, if the previous statement is accurate, this I guess would simply be because the picture was overcompensated for prior to the optimized/more accurate rendering provided by MadVR.

If you compare the various upscaling algos, NGU Sharp is by far the sharpest of the bunch. And even NGU Sharp does *NOT* apply any artificial sharpening. All NGU Sharp does is try to take a good guess at what the original image might have looked like in higher res.

If you already find the source image/video too soft, no upscaling algo will fix that for you. You'll need to use dedicated sharpening algos in that situation.

Originally Posted by ryrynz

Just changed my chroma upscaler from Super-xbr to NGU AA High. Very clean and gives NNEDI3 64+ a run for it's money which were typically pretty pointless basically because of the rendering load and small differences over 32 neurons.

Glad to hear that!

Originally Posted by ryrynz

NGU AA Very High isn't worth it for the rendering time increase based on the comparison I've done (over twice the rendering time on my 960) It seems to have a different profile to the other NGU versions resulting in a fairly minor change & not necessarily improved changes. madshi perhaps you could work something better suited for chroma as a very high setting?

I've mentioned it before: At some point in the future I might look into creating a dedicated NGU chroma algorithm which also takes the luma channel into account, as a guidance for chroma upscaling, similar to what Bilateral or Chroma Reconstruction do. But developing such an algo costs a lot of time, so it will probably not come soon.

Originally Posted by enphenate

My MPC-HC setup has become extremely unstable after upgrading from 0.91.5 to 0.91.7.

Whenever I start a new video and double click it to make it fullscreen, the video glitches out, freezes and becomes unresponsive. The audio will keep playing, then cuts out. If any windows prompt (UAC) opens up, my entire screen goes black and the mouse cursor just flashes really quickly. Can't do anything at this point, the computer is not responsive to any input, the mouse moves around but can't control+alt+delete or do anything. Only a hard shutdown allows me to use the computer again. This only happens when I attempt to make a video full screen. The video is stable when its not in fullscreen. A couple of times this would trigger the MadVR bug report app, which I used to submit a report.

I noticed there are a ton of new settings in this build, may have to redo some of my settings, but I cant imagine what would be causing this since ive never had this issue before.

I tried upgrading MPC-HC to 1.7.11, LAV to 0.69.0 and to the latest Nvidia drivers and nothing has fixed this issue.

Looking at the madVR source code changes between v0.91.5 and v0.91.7, there's nothing in there that would explain this changed behaviour. Are you 100% sure that you didn't change anything else which might explain this change? E.g. a different GPU driver or a different MPC-HC version? Try going back to v0.91.5, I don't think it is likely to fix the issue. I think the issue is most likely to come from a different GPU driver or different media player version.

Originally Posted by dioxholster

Does that mean NGU sharp under image upscaling/doubling or under chroma upscaling?

Image upscaling/doubling.

[direxx]

Does anyone else have issues with 10-bit HEVC playback on AMD video cards?

I have a RX480 and when playing back 10-bit HEVC through MPC-BE (with MadVR) I am getting a black image. I've tried loads of different settings, but could never get an image (only sound).

The weird part is that 8-bit HEVC plays back fine with MadVR. Only 10-bit is affected. My 10-bit HEVC files can be played back by MPC-BE with DXVA directly, without MadVR in between, so I don't think this is a driver issue.

[Neo-XP]

Originally Posted by madshi

I can see that the clock is "interpreted" slightly differently, but I can't really say which interpretation is "better". Anyway, as I said before, NGU Sharp will get some modifications soon. I've no idea if you will like the modifications, though, we'll see...

I can't wait to do an in-depth comparison between the old NGU, the current one and the next one !
For me, the current NGU Sharp (high) lacks sharpness when doing FHD > UHD upscaling, compared to the old one and the current NGU Sharp (very high).

Originally Posted by madshi

I think the test image you're using is not a good one for seeing differences between NNEDI3 and NGU AA/Sharp. In theory it should be because red fonts on black background sound like a good chroma test, but I can barely see any difference. Here's an image where the different is much bigger:

http://madshi.net/chromaTestImage.y4m

Try doubling the image, to make the differences even easier to see.

Still think NNEDI3 is better than NGU AA for chroma?

It is hard to find a good image to test chroma upscaling, but with yours, the result are as follow between NNEDI3 and NGU AA (in order of personal preference) :

Nnedi16 < NGU-low < Nnedi32 < NGU-med < Nnedi64 < Nnedi128 < NGU-high < Nnedi256 < NGU-veryhigh



Quality wise, NGU-high and Nnedi256 are very close, but the differences in performance are huge ! So NGU AA wins for this particular case.
However, I have found NNEDI3 to be better than NGU AA in a lot of situations, for instance to reduce the chroma bleeding.

Originally Posted by madshi

Normally I would flat out say "no" because IMHO luma is *THAT* much more important that a 2 step difference sounds just right to me. However, NGU Med is only slightly slower than NGU Low, which makes me somewhat unsure about this.

Opinions, anyone else?

The problem is that I can't do NGU Sharp (very high) for luma doubling to use NGU Sharp (med) for chroma doubling, I have a lot of dropped frames due to NGU Sharp (very high) only.
However, I don't have any dropped frames with NGU Sharp (high) for luma doubling and NGU Sharp (med) for chroma doubling (tested in v0.91.1 of course), and all the other settings are already maxed out for my GPU.

So it would be awesome if you could allow that

[sneaker_ger]

Originally Posted by direxx

Does anyone else have issues with 10-bit HEVC playback on AMD video cards?

I have a RX480 and when playing back 10-bit HEVC through MPC-BE (with MadVR) I am getting a black image. I've tried loads of different settings, but could never get an image (only sound).

The weird part is that 8-bit HEVC plays back fine with MadVR. Only 10-bit is affected. My 10-bit HEVC files can be played back by MPC-BE with DXVA directly, without MadVR in between, so I don't think this is a driver issue.

Yes, there are several reports of the issue but so far without a real solution except switching to DXVA2 (copy-back). Search the thread.

[leeperry]

Originally Posted by Neo-XP

For me, the current NGU Sharp (high) lacks sharpness when doing FHD > UHD upscaling, compared to the old one and the current NGU Sharp (very high).

Very nice, how about Recon Soft please? That's the real contender in my book, especially with SR2

[Neo-XP]

It does not perform well at all with madshi's chroma test.
It looks all blurry, see for yourself :

[leeperry]

Originally Posted by Neo-XP

It does not perform well at all with madshi's chroma test.
It looks all blurry, see for yourself :

OK thanks a lot, and that's with SR2? Anyway, I've never been a big fan of sharp chroma as if I find it to somewhat interfere with luma too much so I don't think sharpness is the best benchmark to compare chroma upscalers to my eyes........anyway I run a 4:2:2 TV and test patterns are just that but fair enough I'll roll them all all over again on real-world content for OCD sake's ^^

[Neo-XP]

Yes, that is with SR2, you can see it in the stats.

[ryrynz]

Originally Posted by leeperry

I don't think sharpness is the best benchmark to compare chroma upscalers to my eyes.....

I feel the same way. While I've done this to show the differences between them, it's not exactly ideal. I think one should be looking at the naturalness in appearance, cleanness and ringing with brightness just being a matter of taste.

Thanks to NGU AA I think NNEDI3 has been superseded for chroma mostly due to its performance requirements.

[70MM]

Is NGU Sharp different to the first NGU we had?

[leeperry]

Originally Posted by ryrynz

Thanks to NGU AA I think NNEDI3 has been superseded for chroma mostly due to its performance requirements. Brightness is a matter of taste.

Quite frankly I think madshi's NGU-Sharp has buried NNEDI3 for good now, ludicrous PQ

[galneon]

Originally Posted by madshi

If the context menu doesn't appear, trying to right click in the middle of the screen, wait for half a second, then move the mouse a little to the left and top and right click again (blind). Maybe that helps as a workaround?

I suppose at some point it might make sense to somehow detect your right click and disable FSE mode for a couple of seconds, *before* showing the context menu. Right now it's the other way round: The context menu is displayed. Then madVR detects that there's a context menu which won't be visible in FSE mode, so madVR switches to windowed mode, but depending on OS, GPU drivers and media player, the switch to windowed mode may make the context menu close again...

I enabled FSE again to attempt your workaround and was surprised to find FSE switching to windowless for the context menu (and then back again) was back to working as usual even without a workaround. All that has changed is my video driver, from nvidia 376.xx to 378.78. I don't know why that would be the determining factor, but if I'm positive it's the only thing that has changed, then I suppose it was. Thanks for giving me a solution to try had the problem not worked itself out.

[QBhd]

Originally Posted by madshi

Normally I would flat out say "no" because IMHO luma is *THAT* much more important that a 2 step difference sounds just right to me. However, NGU Med is only slightly slower than NGU Low, which makes me somewhat unsure about this.

Opinions, anyone else?

There are a few scenarios where I can't push the Luma higher and have room left for more Chroma, but with the limitations currently in place I end up leaving GPU power on the table. I know where you are coming from with these limitations, but I for one would be happy if they were lifted.

QB

[ryrynz]

Originally Posted by madshi

What do you mean with "resets back to the stop"? I'm not sure.

Typo, I meant top. Every time I tried to enter something the scroll bar reset making me need to scroll back down again.

Originally Posted by madshi

So you're saying you're prefering NGU AA over NNEDI3 now? If so, I'd be happy to hear that.

The one thing I disliked about NNEDI3 was it's slightly excessive sharpness in areas that didn't really warrant it. Since I already have two sharpeners in my process (Awarp and AS) now I find these together produce a more balanced image.
Where I thought sharpness was beneficial in stills while watching I realized it wasn't actually what I preferred.

Originally Posted by madshi

Yes, something in between high and very high is planned for the future, but it all costs serious development time, so I can only deliver things step by step.

Glad to hear it.

Originally Posted by madshi

Right now all the NGU algos are not perfectly fine tuned yet. I'm still doing rough adjustments right now. Once all the rough coordinates are dialed in, I'll fine tune all the algos and presets one by one. But that will take time, so don't expect it any time soon.

Great, something to look forward to.

Originally Posted by madshi

Maybe we need to collect better test images/samples for chroma upscaling. If I want to judge which algo is better, a test image should show something which jumps right into your eye.

I think most of us just want accuracy, we need someone to deliver some 4:4:4 ground truth (actual video) and convert it to 4:2:0 for some comparison, I am a bit surprised nobody has done this yet.
Anyway, TBH NGU AA High is very nice for chroma IMO with medium and even low being good options too. Honestly given the performance and results of these I'd rather time was spent elsewhere..

perhaps a built in replacement for ffdshow raw? Being able to add in Avisynth+/Vapoursynth filters into madVR directly would allow for huge video customization and allow also for proper handling of 10 bit content (ATM I have to disable it for any of this video type to avoid 8 bit output) and hopefully be faster and more stable as a result.

[andybkma]

madshi, subtitle bug to report. Win7, mVR 91.4, xySubFilter 3.1.0.746. I have mVR set to move the subs to the bottom of screen/window. Never had a problem until last night. Starting the movie was fine and the standard font subs displayed fine under the screen/window as they should per mVR settings. But then italics subs appeared being used to show dialogue from a character off screen and those italicized subs displayed in the window instead of below it. From that point on all subs were then displayed in the window whether italics or standard font, so mVR was now totally ignoring instructions to display the subs below the screen. Stopping media player then starting it again fixed the problem until the next time italics subs appeared then all were displayed back in the window again. Seems mVR does not like italics subs or switching from standard font to italics causes it to wig out?

The file is 7GB so uploading a sample with matching SRT subs would seem to be impossible... and not sure if this bug reported yet. Thanks much

[YxP]

Originally Posted by andybkma

not sure if this bug reported yet. Thanks much

It's a known bug in xysubfilter, but the author seems to have disappeared.

[Telion]

And this bug appears only when SRT subs are in an external file, but not when they are muxed into an MKV file. So one can avoid it by just remuxing a video with subs into a single container, though remuxing huge videos just for that purpose can be PITA, of course.

[cyber201]

Hi,
It's normal when I play an HDR video with Windows10 and KODI or MPC-HC, the madvr show me this info:

matrix BT.2020
primaries BT.2020
HDR 4000 nits, BT.2020 -> DCI-P3

Why DCI-P3 and not BT.2020?

This is for all HDR videos

thanks a lot

[CruNcher]

Because only expensive Devices can actually display the Full Bt 2020 space that are no option yet for consumers and inside Research Labs.

DCI-P3 is the most economical solution currently for the masses without tonemapping the entire result to something different, even payable expensive devices only emulate Bt 2020 to some degree.

It will take some years and solds before Bt 2020 becomes reachable to normal consumers, some very long years when we predict it takes the same time as 10 bit finally reached end consumers or full bt 709 and also actually become payable to run them for a long time

[HelmedHorror]

What might be causing dropped frames even when the render time is like 10ms for 24fps content? I understand dropped frames when the render time in seconds is greater than the reciprocal of the fps, but that's not the case here.