madVR - high quality video renderer (GPU assisted)

[vivan]

Originally Posted by James Freeman

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

LAV Filter should dither only if output colorspace is not the same as input (video). That means when it's not Y416 for that video.
[s]However that dithering should be in lsb, so if you dither 16 to 10 and then truncate it to 8 it should be the same as truncate it from 16 to 8 directly.[/s]
UPD: that's not true. If you have, for example, 1000000 pixel and negative error it will be changed to 01111111...

Originally Posted by Ver Greeneyes

Yes, although I'm not using hardware decoding.

You can't use hardware decoding for anything other than High/Main/Baseline AVC profiles (means 8-bit 4:2:0 only) anyway.

[James Freeman]

Originally Posted by Ver Greeneyes

Yes, although I'm not using hardware decoding.

Hardware (CUVID, DXVA2) or Software (None), all the same forced dithering in Y416 mode.

I'm on HDMI right now maybe its related?
I'll switch to DisplayPort and report.

EDIT:
No, same forced dithering on DP with Y416.

EDIT: Fixed, DeInterlacing was On.

[Ver Greeneyes]

Originally Posted by James Freeman

I'm on HDMI right now maybe its related?
I'll switch to DisplayPort and report.

I wouldn't expect it to make a difference - madVR outputs 8-bit per color channel regardless of the input color space.

Originally Posted by James Freeman

Please generate the same videos (Grey and Color) at 16-bit 10-bit & 8-bit (4:2:0, 4:2:2 & 4:4:4).

Should the input be RGB or YUV? My program currently outputs RGB48LE.

[mithra66]

A1 was a chock when Madshi released it last night. You do such a great job , Madshi!

I Had to wait for night as I couldn't tell the difference even with minimum daylight.

So I prefer A1 > A3 >= A2.

I agree that A1 is darker as if gamma was slightly raised. But it has no inconvenience such as oversaturated color or so on my display. On the contrary.

[James Freeman]

Originally Posted by Ver Greeneyes

Should the input be RGB or YUV? My program currently outputs RGB48LE.

Same as the current videos.

I'll try to explain what I see and what I'm trying to test:

When I disable Y416, the decoder goes to the closes lower next one: v410.
When v410 is disabled it goes to Y410, and so on till YV24.
Each of these levels has DIFFERENT forced dithering pattern !!!

When all the 4:4:4 row is disabled it goes to the 16-bit RGB -> RGB48 which is not dithered at all.
That is what we are striving for, to let MadVR do the heavy lifting.

I want to test if this forced dithering is somehow related to Nvidia or LAV filters with various formats of video.
These videos should be extremely useful to all video testing, not only this case.

I'll post all the different variants of forced dithering that I see with different output formats.

EDIT: Fixed, DeInterlacing was On.

[madshi]

James, does this have anything to do with madVR? If you want to analyze something about how LAV or NVidia do dithering, then this is not really the right thread for that.

[James Freeman]

Sorry for that madshi.

EDIT:
Nevermind, Solved.

It was the Automatic On DeInterlacing that dithered the 16-bit image.

[Ver Greeneyes]

Originally Posted by James Freeman

Same as the current videos.

Let me put it in the terms ffmpeg uses. There are YUV444, YUV422 and YUV420 variants for 8-bit, 10-bit and 16-bit. In total there are 9 possible combinations. Which of these do you want?

yuv420p
yuv422p
yuv444p
yuv420p10le
yuv422p10le
yuv444p10le
yuv420p16le
yuv422p16le
yuv444p16le

Edit: I'll move this to the new thread as well.

[James Freeman]

Ver Greeneyes,

I had DeInterlacing turned on automatically.
This what was dithering the picture.

Vivan was on to it as soon as I posted the first image.

[NicolasRobidoux]

Originally Posted by madshi

Hmmmm... I like madVR's Jinc3 AR better. And NNEDI3 is another class, IMHO...

Note that except for the conversion to linear RGB (which I'm not sure is the best choice with video content when upsampling, but may be with pixel art) what I showed is a purely linear scheme: There is no AR applied whatsoever.

AR is your department. (Until I program my own; don't hold your breath )

In any case, I tracked down the general purpose scheme I like with pixel art. This one: http://www.wizards-toolkit.org/disco...tart=30#p78393

Result of running

convert SNES.png -define filter:lobes=4 -define filter:blur=0.88451002338585141 -filter Lanczos -distort Resize 1234x lanczosSharpest4.sRGB.png

http://web.cs.laurentian.ca/nrobidou...throughRGB.png

Whether, with AR, it satisfies you or others, let alone beats NNEDI, is another story. Also, maybe it should only be applied to luma. But note how noticeably less jaggy it is on perfect diagonals, how tight the halo is, and how sharp it is, compared to the 3-lobe version I showed earlier. Also note that the corresponding disc is barely larger than EWA LanczosSharp, because the deblur is so strong. So it should run reasonably fast.

A slightly less sharp version, EWA LanczosRadius 4 (Jinc-windowed Jinc 4-lobe with support shrunk so it is a disc of radius 4):

convert SNES.png -colorspace RGB -define filter:lobes=4 -filter LanczosRadius -distort Resize 1234x -colorspace sRGB SNES.lanczosRadius4.throughRGB.png

http://web.cs.laurentian.ca/nrobidou...throughRGB.png

The jist is this: If you deblur EWA Lanczos (Jinc-windowed Jinc) a lot, and you want to do well with thin and artificially sharp diagonals (hello pixel art), 4 lobes is better than 3 lobes. (For natural images, I generally prefer 3 lobes and 5 lobes.)

Yes, these EWA LanczosSharpest/LanczosRadius don't antialias as strongly as, say, EWA LanczosSharp (one version of which is your Jinc3 without AR). But they are tight.

NNTR

[Ver Greeneyes]

Originally Posted by James Freeman

I had DeInterlacing turned on automatically.
This what was dithering the picture.

Mystery solved!

[wolfman2791]

Originally Posted by James Freeman

Ver Greeneyes,

I had DeInterlacing turned on automatically.
This what was dithering the picture.

Vivan was on to it as soon as I posted the first image.

So i'm assuming the "Auto" option is not what you want in LAV Video settings? So what is the setting that you want??

[James Freeman]

Auto is fine in LAV & MadVR,
but set MadVR to "if in doubt, DE-ACTIVATE deinterlacing".

[leeperry]

Originally Posted by madshi

And what is your opinion of how A3 compares to NL6?

Well, kinda tricky question as all non-A builds look obsolete now due to the excessive grain but A2 & A3 are wasted potential and a waste of time once you've seen A1 IMHO.

Basically with A1 every moving object has its edges clearly defined but with A2/A3 it's just a big blurry mess that kills the depth impression(call it "3D" if you have to) and the picture dynamics...even RD looks better to my eyes than this big foggy mess.

Originally Posted by mithra66

A1 was a shock when Madshi released it last night. You do such a great job , Madshi!
[..]
I prefer A1 > A3 >= A2.

Wholeheartedly agreed, A1 is da bomb

[GCRaistlin]

Where can I download old versions of madVR (<0.86.1)?
It's about the issue I've reported above. MPC-HC developer has confirmed the issue with the latest madVR, ffdshow and MPC-HC but doubts that it's MPC-HC related bug. On the other hand he said that it was not reproducable with the older madVR and/or ffdshow builds though he has had overwritten both of them with the latest versions and he doesn't remember old version numbers. I've tried a lot of versions of ffdshow with madVR 0.86.1, 0.87.3 and 0.87.4 - MPC-HC (as well as MPC-BE) hangs on mouse moving over buttons in the sample DVD root menu.

[James Freeman]

Grey:

A2 is the closest to NL6 in terms of gamma, but it has better blacks.
A1 is similar to A2 but is a little brighter at the middle.
A3 is the brightest in the middle and has similar blacks as A1 and A2.

In terms of Gamma A2=NL6 (darkest or perfect?) > A1 > A3 (brightest).
Also, NL6 has more "holes" (undithered space) compared to the Adaptive builds.

A2 Wins.

Color:

A1 looks too fluffy, not dense enough.
A2 Nice and even.
A3 = A2, look exactly the same.
NL6 looks fluffy like A1 but with cut blacks.

A2, A3 Tie.


Overall A2 has the Positives from NL6 without the Negatives.

[Devrim]

What are the recommended settings with the new update?

Atm I have Jinc3 AR for Chroma and Image upscaling en C-R with AR en SLL for image downscaling but what about NNEDI?
How can we use that?

Also I noticed some banding on a BR I was watching, the artifac removal took care of that, beautifull feature! I set it to high.
Can I leave it at that? Should I set it to low/medium? Or just turn it off and only turn it on when I see banded image?

[Shiandow]

In my opinion artefact removal should be as low as possible, since it can remove some details. I use the default settings and use a hotkey to set it higher when I see banding.

[Devrim]

Originally Posted by Shiandow

In my opinion artefact removal should be as low as possible, since it can remove some details. I use the default settings and use a hotkey to set it higher when I see banding.

http://screenshotcomparison.com/comparison/62937
http://screenshotcomparison.com/comparison/62938
http://screenshotcomparison.com/comparison/62939

There are only small differences. (Sorry for the not sharp screenshots with not much detail)

[6233638]

Originally Posted by NicolasRobidoux

In any case, I tracked down the general purpose scheme I like with pixel art.
...
http://web.cs.laurentian.ca/nrobidou...throughRGB.png
http://web.cs.laurentian.ca/nrobidou...throughRGB.png

Honestly, pixel art does not look good when you use video-style filters on it. It's just not how it was intended to look.

Something like this is more appropriate for gaming. (it is expected that you turn up your display brightness to compensate for the scanlines)


We really need much higher resolution displays for proper CRT emulation, but some of the approximations are getting pretty good (to the point that I would actually prefer to use them rather than avoid them) and there are probably much better filters now than the one I used for that image.

P.S. I hate the vignetting and rounded-off corners seen here. Some aspects of CRTs are worth emulating (scanlines, beam width, pixel structure etc.) others are not. (phosphor persistence, geometric distortion, vignetting, floating black level/low ANSI contrast, composite video connectors etc.)

But this is getting quite off-topic...