madVR - high quality video renderer (GPU assisted)

[6233638]

FSE mode throws off sync for me, so I disabled it. The lips don't match 100% with the audio, it's slightly off. Enough to be annoying. Turning FSE mode off fixes the issue, so I've abandoned it until it's addressed.Try the "delay playback start" rendering options.

[James Freeman]

6233638, Ver Greeneyes, bacondither

How do you create this effect to clearly see the dots (dithering), so I can see it first hand too.
What soft?

1. Take a screen shot.
2. ...

[iSunrise]

6233638, Ver Greeneyes, bacondither

How do you create this effect to clearly see the dots (dithering), so I can see it first hand too.
What soft?

1. Take a screen shot.
2. ...
1) You load the untouched images or the clips into madVR, skip to the same frame (if it´s a clip), go into windowed fullscreen, take a screenshot, use an image viewer/editor that is able to alter the input level range, adjust the max. level (should be at 255 originally) to something that shows you all bars or even lower to make lower steps more visible and save the image. I use GIMP for this (it´s free).

2) You basically do 1) with every new build, then you can compare (I use Irfanview for this, because it has a nice slideshow function that will make comparisons very easy) the end result.

[James Freeman]

Thanks iSunrise.

Yes I tweak the Input/Output levels in Paint.NET (also free and very compatible).
But still, I can't get that single bit dotted image like you guys post to compare the dithering versions.

[ryrynz]

And NNEDI3 is another class, IMHO.
http://screenshotcomparison.com/comparison/59698


If only it didn't ruin the color and size of the x's and what's with all those vertical artifacts around the x's the numbers and letters?
I would love to see this somehow addressed, more neurons? (maybe not) or slightly better detection of things that should be left alone, even if it meant a loss of sharpness..

I don't know if NNEDI will see further development, It does so much so well, and fails hard in the rest.. I wish there was a way to minimize the problem areas, which I was hoping resizing before applying NNEDI might achieve
but that seems to kill much of the benefits of NNEDI.

[iSunrise]

If only it didn't ruin the color and size of the x's and what's with all those vertical artifacts around the x's the numbers and letters?
I would love to see this somehow addressed, more neurons? (maybe not) or slightly better detection of things that should be left alone, even if it meant a loss of sharpness..

I don't know if NNEDI will see further development, It does so much so well, and fails hard in the rest.. I wish there was a way to minimize the problem areas, which I was hoping resizing before applying NNEDI might achieve
but that seems to kill much of the benefits of NNEDI.
I already mentioned this in a post of mine a couple of pages ago, along with an image of certain problem areas that look like ringing artifacts to me. An additional pass through madshi´s AR filter could probably eliminate almost all of them (if these are indeed added through NNEDI3 and aren´t in the source itself), but I´m not sure if that would introduce other artifacts. Since there´s no way for us to test that, I just wanted it to get out in the open as a thought.

Thanks iSunrise.

Yes I tweak the Input/Output levels in Paint.NET (also free and very compatible).
But still, I can't get that single bit dotted image like you guys post to compare the dithering versions.
Here (http://forum.doom9.org/showthread.php?p=1668251#post1668251) is the source (expanded original) that everyone seems to be using. Just do what I described above on it and it will work, it does for me.

[cyberbeing]

Took a quick look at the latest builds.

Adaptive2 had the least banding on near-black and near-white gradients.
NL6-Gamma had the least banding on mid-tone gradients.

[madshi]

Does madvr support 64-bit MPC BE or HC yet?
No.

FSE mode throws off sync for me, so I disabled it. The lips don't match 100% with the audio, it's slightly off.
That shouldn't happen. Seems most other users don't have this problem, so it's probably specific to your hardware or configuration. You could try lowering the size of the GPU queue and/or the number of frames that are presented in advance. Some GPU drivers don't like it if these numbers are too high.

Humm, that's plain ridiculous. The AGL1 build looks so much better than NL6.....need to test again tomorrow morning but the latter looks unnecessarily grainy and unfocused compared to the AGL1 build. The LL builds still look grainy and plain ugly, I don't see how that could change with fresh eyes but we'll see :)
Well, I suppose I should call myself a "scientist" then. Anyway based on the images 6233638 posted I'd say that adaptive is better. In an effort to make this somewhat more objective I also tried blurring the images. This confirmed that adaptive gave the 'smoothest' gradient and that unlimited was both more noisy and too dark.
Ok, thanks for the feedback.

Took a quick look at the latest builds.

Adaptive2 had the least banding on near-black and near-white gradients.
NL6-Gamma had the least banding on mid-tone gradients.
And how does the first adaptive build fit into this? Which of three do you prefer?

[iSunrise]

Just FYI, for anyone that has problems getting still images to show in madVR (black screen), you absolutely need to enable "delay playback start until render queue is full", otherwise madVR seems to have a problem displaying certain images (in my case, several 16bit/8bit per component TIFF and PNG files). I´m not sure why it works for some, but not for others. Ver Greeneyes' PNG samples (posted the link above) show the problem, while other files work fine. I don´t quite get it.

madshi, is that a bug? If it is, I´m going to add it to the bug tracker right away.

[Ziron]

If only it didn't ruin the color and size of the x's and what's with all those vertical artifacts around the x's the numbers and letters?
I would love to see this somehow addressed, more neurons? (maybe not) or slightly better detection of things that should be left alone, even if it meant a loss of sharpness..

I don't know if NNEDI will see further development, It does so much so well, and fails hard in the rest.. I wish there was a way to minimize the problem areas, which I was hoping resizing before applying NNEDI might achieve
but that seems to kill much of the benefits of NNEDI.

I really wish tritical had open sourced his training suite for NNEDI. It would be interesting to see if training it on more/different data would produce less/different artifacts. Given the renewed interest in NNEDI recently though we might get lucky and have someone reverse engineer some new neural network training software for it.

[cyberbeing]

And how does the first adaptive build fit into this? Which of three do you prefer?

Well my testing was limited in scope to various linear gradient patterns only, but overall I preferred Adaptive2.

Adaptive1 had around the same amount of near-black and near-white banding as NL6-Gamma, and both Adaptive builds were slightly worse than NL6-Gamma on mid-tones. Overall Adaptive1 seemed to have the most banding out the the three.

[DragonQ]

There's a great deal on an AMD R9 270X (2 GB GDDR5) at the moment. Can someone with this or a similar card comment on what it's capable of in terms of MadVR settings?

[iSunrise]

Well my testing was limited in scope to various linear gradient patterns only, but overall I preferred Adaptive2.

Adaptive1 had around the same amount of near-black and near-white banding as NL6-Gamma, and both Adaptive builds were slightly worse than NL6-Gamma on mid-tones.
Overall I can agree with this, the adaptive builds seemed a bit darker to me overall (which doesn´t necessarily mean it´s bad) and they look a tiny bit more saturated in the overall image appearance (skin tones), which seems to be a side-effect of that.

The NL6-GL build seems to be a little on the brighter side, since it shows a bit more details in dark areas. Probably related to the dot placement/dot intensity.

[bacondither]

Adaptive 1 (http://s27.postimg.org/o6h9gzyg3/adaptive1.png)

Adaptive 2 (http://s27.postimg.org/fa6hd27tv/adaptive2.png)

----------------------

Adaptive 1 Blurred (http://s9.postimg.org/a7l4grui7/adaptive1blur.png)

Adaptive 2 Blurred (http://s9.postimg.org/phkznyq0f/adaptive2blur.png)

----------------------

I prefer Adaptive 1.

[James Freeman]

bacondither,
Would you be kind to tell me how you generate those patterns?

iSunrisem
Your method did not work, it seems to clip everything except the few lines that left.
But it does not look anything like what bacondither, cyberbeing or Ver Greeneyes posting.

[iSunrise]

I prefer Adaptive 1.
Can you give an explanation, why exactly you prefer it?

The only thing that I really see different from your pics, is that Adaptive 2 has a less random dot placement around the "edges" of the bars. And since the bars themselves are identical from top to bottom, the dots should also be less random, they should be more aligned IMHO.

With bar comparisons like this, that don´t show us anything apart from the dither dot placement, an algorithm should behave in a more expected manner, instead of a more random manner, also IMHO.

But that´s free to discussion of course.

[bacondither]

Can you give an explanation, why exactly you prefer it?

If you look at the blurred adaptive 2 it has a tiny bit more banding then then blurred adaptive 1, but besides that they are almost identical.

[cyberbeing]

The dither pattern of Adaptive2 is much smoother than Adaptive1:
http://screenshotcomparison.com/comparison/62852

[bacondither]

The dither pattern of Adaptive2 is much smoother than Adaptive1.

What pattern/file are you using?

[leeperry]

Just spent quite a lot of time comparing all the newest builds on 1080p24 and 720p60 content:
-All builds look grainy and unfocused compared to the new adaptive ones, the nice side is that the grain hides upscaling/compression artifacts as I was forced to disable NNEDI for 720p60 and J3AR is a major step backwards once all that covering grain is gone using AGL1
-NL6 provides a deeper pop-effect than either AGL builds but it's fatiguing to look at as my brain has to dig behind all the artificial grain trying to reassemble all the details, the picture is far more natural and clearer looking with AGL1
-AGL1 looks more transparent and much clearer than AGL2, the latter is not nearly as versatile and it looks like a veil that kills dynamics and sharpness to my eyes, I even prefer random dithering to AGL2...otoh AGL1 is a huge improvement over RD.

[cyberbeing]

What pattern/file are you using?

That one was from Ver's gradient-perceptual-v2.mkv:


Frame 250
madVR Calibration Disabled
madVR Custom Debanding (avgDif = 0.3, maxDif = 1.3, midDif = 0.6, angleBoost = 2.0, maxAngle = 0.08)
Saturation boosted in Photoshop to show dither pattern

[Shiandow]

If you look at the blurred adaptive 2 it has a tiny bit more banding then then blurred adaptive 1, but besides that they are almost identical.

I agree although the difference is very small.

[bacondither]

Here is some images from "colored gradient.mkv" from Ver Greeneyes, frame 88 and histogram clipped at 6.

Adaptive 1 color (http://s28.postimg.org/qynynowtp/adaptive1color.png)

Adaptive 2 color (http://s28.postimg.org/wy74dls8d/adaptive2color.png)

Blurred using gimp 2.9 in linear gamma then converted back to gamma space.

Adaptive 1 color blurred (http://s27.postimg.org/cn7pxy1pf/adaptive1colorblur.png)

Adaptive 2 color blurred (http://s27.postimg.org/rkg6yyexv/adaptive2colorblur.png)

-------------------------

Here i think adaptive 2 looks marginally better... oh the ocd level is too high... :sly:

[Shiandow]

[...]

Here i think adaptive 2 looks marginally better... oh the ocd level is too high... :sly:

They look identical too me. By the way if I blur them myself, in gamma light not in linear light, then I see even less banding than in your images; it's possible that the converting from linear light to gamma light caused some of the banding.

[James Freeman]

Can anyone tell me how you create those comparison images? Anyone?

There are basically 3-4 people who can really compare the test builds, everybody else is just on standby.

[bacondither]

They look identical too me. By the way if I blur them myself, in gamma light not in linear light, then I see even less banding than in your images; it's possible that the converting from linear light to gamma light caused some of the banding.

Blurring in gamma light will give you the wrong results! Wrong as in (5+70+20)/3=40.

Conversion from linear light->blur->gamma light is done in 32-bit FP.
I am not dithering down when going back to 8-bit(mabe i should?) and it should introduce the same banding as doing it the wrong way i.e blurring in gamma light.

[Shiandow]

Blurring in gamma light will give you the wrong results! Wrong as in (5+70+20)/3=40.

Conversion from linear light->blur->gamma light is done in 32-bit FP.
I am not dithering down when going back to 8-bit(mabe i should?) and it should introduce the same banding as doing it the wrong way i.e blurring in gamma light.

Well technically it is correct if you use a sRGB gamma curve, since this will be linear for values that dark. But I mainly pointed it out because your method seemed to introduce more banding, which makes both algorithms look worse than they actually are.

Edit: Anyway I redid the blurring using linear light, which gave the following results:

Adaptive 1 blurred (http://postimg.org/image/4h4mva0ot/)
Adaptive 2 blurred (http://postimg.org/image/j1lpq3vnh/)

these seem to have less banding than the ones in your post.

[Shiandow]

Can anyone tell me how you create those comparison images? Anyone?

There are basically 3-4 people who can really compare the test builds, everybody else is just on standby.

You need some kind of image processing software, anything more advanced than paint should have an option somewhere to brighten the images.

[James Freeman]

I use Paint.NET (like Gimp) which has all the manipulation options that needed.

I tried to limit the Input Levels from 0-255 to 0-6, but no good, it clips everything above 6 and leaves several thin lines in the most dark shades.
Do you know what people here do to create the results they get?

P.S
I'm pretty sure I'm being ignored for being an asshole (or just having an opposite opinion) few pages back, but never mind, I'll continue to participate in this magnificent project, ignored or not.

[madshi]

It's interesting. It seems there's no clear consensus here. Some prefer Adaptive1, some Adaptive2. Maybe we should try to find a build which everybody likes? Here comes the next try:

http://madshi.net/madVRadaptive3.rar

This one should be quite similar to the original NL6, but should have less banding with certain gray levels. So I hope it's a clear improvement over NL6 without any disadvantages. What do you think?

[bacondither]

Adaptive build 3 removes some minor patterns between interger levels(0-255).

Looks good, great work madshi! :)

[Ver Greeneyes]

Can anyone tell me how you create those comparison images?
0) In the madVR settings, disable calibration and disable debanding (debanding might not matter, but it caused some odd effects on the colored version of my gradient at least)
1) Get my perceptual gradient (http://www.mediafire.com/?eh01w3vrfshk5oe)
2) In MPC-HC, go to the Navigate menu and select Go To..., then go to a particular frame you want to use for testing. Use a dark frame so you can brighten it without clipping.
3) Take a screenshot with the Print screen key on your keyboard
4) Open up an image editor like GIMP or Photoshop and paste the screenshot from the clipboard
5) Crop the image to just the 1920x1080 video
6a) Using GIMP, go to Tools -> Color Tools -> Levels..., then move the rightmost arrow below the Input Levels left until the brightest shade in the image is white, and press OK.
6b) Using Photoshop CS6, go to Image -> Adjustments -> Levels..., then move the rightmost arrow below Input Levels left until the brightest shade in the image is white, and press OK.

Note the brightness may differ slightly between dithering variants, so you may want to wait until you have screenshots of each version before you start messing with the levels, then choose the maximum input level to be the same for each of them.

And finally, I used a custom program for my latest comparison (http://forum.doom9.org/showthread.php?p=1668270#post1668270) (already out of date compared to the latest builds), which uses a conversion into the perceptual CAM02-UCS space to brighten the image. As you can see from the high amount of saturation in the images, this is not without its difficulties.

[James Freeman]

Thanks Ver Greeneyes. :)

But, The results with your pattern are:
http://www.mediafire.com/convkey/f134/a6hgmg7xug56gt3fg.jpg

This is with ED Adaptive 2.
Even with dithering completely disabled it is the same.

Why is it colorful?

EDIT: Fixed, DeInterlacing was On.

[nevcairiel]

Why is it colorful?

Maybe you have random dithering active?

[Ver Greeneyes]

Why is it colorful?
This looks like random dithering. Can you check that your decoder is passing through 16-bit color formats? For instance, I have the LAV Video Decoder set up like this: settings (http://i.imgur.com/fEKvGgQ.png). MPC-HC seems to automatically fall back to the standalone LAV filters for my video for some reason by the way, and crashes with its internal version. I actually had the standalone LAV filters configured to only output RGB24 to appease a game, so initially I wasn't viewing my own test pattern properly XD

[James Freeman]

Maybe you have random dithering active?

Nope. I have it completely disabled.

I think the original video contains dithering.
Just by turning brightness in MPC-HC I see the colorful dithering.

EDIT:

Ver Greeneyes,

LAV Video filter 0.60.1.22
Same settings as yours.
Same colorful results.

It really looks like the video is natively dithered.

EDIT: Fixed, DeInterlacing was On.

[Ver Greeneyes]

I think the original video contains dithering.It doesn't.

Same settings as yours.
Same colorful results.There must be a configuration error somewhere. If you have dithering disabled in madVR, then something else in your processing chain is introducing it.

Edit: If you right click the window and go to the Filters menu, does it show anything other than madVR, LAV Video Decoder and LAV Splitter Source (internal)? (I also have the XySubFilter beta, but that's not causing issues)

[leeperry]

A3 is a slight improvement over A2 sharpness-wise but A1 is much clearer and transparent than either of them to my eyes.

[James Freeman]

It doesn't.

There must be a configuration error somewhere. If you have dithering disabled in madVR, then something else in your processing chain is introducing it.

Edit: If you right click the window and go to the Filters menu, does it show anything other than madVR, LAV Video Decoder and LAV Splitter Source (internal)? (I also have the XySubFilter beta, but that's not causing issues)

No additional filters.

Changing the renderer back to EVR, and the colors are gone.
It seems MadVR always dithering the image with 16-bit video on my system.
The videos from AVS709HD are OK (I can disable/enable ED or RD).

EDIT: Fixed, DeInterlacing was On.

[noee]

@James Freeman, make sure the device you're playing back is set as 8-bit in madVR device settings.

[James Freeman]

@James Freeman, make sure the device you're playing back is set as 8-bit in madVR device settings.

It is. Always has been.

Changing the brightness in MPC or MadVR to maximum should not have an effect on the dithering right?
But it does with Ver Greeneyes video.

That what makes me think its in the video.

All the "Trade Quality For Performance" settings are disabled.

EDIT:
Double sizing the video in MPC double sizes the dithering noise.
Its definitely in the video.

EDIT: Fixed, DeInterlacing was On.

[madshi]

A3 is a slight improvement over A2 sharpness-wise but A1 is much clearer and transparent than either of them to my eyes.
And what is your opinion of how A3 compares to NL6?

[nevcairiel]

That what makes me think its in the video.


The video is fine, looks perfect here.

[James Freeman]

The video is fine, looks perfect here.


Does it show colorful dithering when Brightness is set to 100 in MPC-HC? or is it perfectly black & White?
MadVR reports Y416, 16-bit, 4:4:4.

EDIT: Fixed, DeInterlacing was On.

[James Freeman]

OK, I solved the dithering in 16/10 bit issue.

I disabled Y416, v410, Y410, in LAV Video, they automatically turned on some bizarre colorful (non MadVR) dithering.
Then I disabled AYUV, YV24 which also dithered the image but in black & white.
I disabled all the 4:4:4 line in LAV Video Filter so now the image renders as RGB48LE.
Its perfect now.

Now I see great banding when the video plays unless I enable MadVR dithering.
Can anyone confirm that that is how they see it?

EDIT: Fixed, DeInterlacing was On.

[vivan]

Now I see great banding when the video plays unless I enable MadVR dithering.
Can anyone confirm that that is how they see it?That banding is the reason why you should always use dithering ;)

[James Freeman]

That banding is the reason why you should always use dithering ;)

Thanks vivan, I know. ;)

IMO, Nvidia thought they doing us a favor by automatically dithering 16/10 bit content.
Or is it something LAV Filter does?

Can anyone please check what MadVR uses for rendering Ver Greeneyes's video (Y416, v410, Y410, AYUV, YV24).

EDIT: Fixed, DeInterlacing was On.

[Ver Greeneyes]

Can anyone please check what MadVR uses for rendering Ver Greeneyes's video (Y416, v410, Y410, AYUV, YV24).It uses Y416 here.

[James Freeman]

It uses Y416 here.

Nvidia?

With Y416 I get extremely noisy forced rgb dithering.

EDIT: Fixed, DeInterlacing was On.

[Ver Greeneyes]

Nvidia?

With Y416 I get extremely noisy forced rgb dithering.Yes, although I'm not using hardware decoding.