madVR - high quality video renderer (GPU assisted)

[wolfman2791]

Originally Posted by michkrol

If you mean "The Intel Media SDK Software MPEG2, VC-1 and h264 Decoder DLL" from first post, you only need it to use madVR's internal decoding capabilities instead of your codecs (LAVFilters/ffdshow/?). It's not recommended anymore, better use LAVFilters, so generally you don't need them.

Thanks. So it's bypassed if you have LAV Filters turned on and it's used as a default if you don't... basically not needed.

[Qaq]

Originally Posted by turbojet

Forcing video works but leaves a duplicate every sixth frame.

Forcing "video"? You had to force "Movie" to let madVR IVTC it 29>23.

[madshi]

Originally Posted by Anime Viewer

Is it still the case that your DirectCompute NNEDI3 tests show it to be 10 times slower than with OpenCL?

Would you elaborate on what was slower? Are you referring to the rendering ms speed? What was the hardware (Intel, AMD, Nvivida), and Madvr settings (including downscaling choice)?

Perhaps a different combination of settings would work better with the DirectCompute coding compared to OpenCL, or perhaps results would differ when tested with an Nvidia equipped system compared to an AMD.

I've not spent any further time on that yet. Overall rendering time was 10x slower. Tested on AMD, but I'm pretty sure that the same problem would occur with NVidia, too, cause the DirectCompute compiler comes from Microsoft, AFAIK. The GPU drivers just run an "optimizer" pass after that.

Originally Posted by Anime Viewer

With the extreme amount of bugginess the Nvidia systems have with the OpenCL coding it may still be better off (even after Nvidia patches the problem) for Nvidia systems to use the DirectCompute version.

Would why any NVidia user choose to use a 10x slower version? Makes no sense to me. Anyway, I've no interest in discussing this now, I've not completed development yet, was just giving you guys a heads up on current status.

Originally Posted by XMonarchY

1. Is there a way to render JPEG files with madVR to improve/sharpen them as if they were just single film frames?

I think you can just drag them into MPC-HC? I don't know, this kind of stuff is the job of media player or splitter/source filter developers. If they support JPEG files, madVR will happily render them.

Originally Posted by XMonarchY

2. My Samsung LN40B550 supports full 4:4:4 only in 60Hz mode, which is not as smooth (even with Smooth Motion) as the 48Hz 4:2:2 judder-free mode that I prefer. However, the 48Hz mode has 2 versions - Movie and Standard. Movie 4:2:2 mode creates nasty text discoloration and on your 4:2:2/4:4:4 test image I can see 4:2:2 clearly, but I can also see 4:4:4 with similar nasty/ugly discolorated edges around it. Then there is the Standard 4:2:2 mode that I use because unlike Movie 4:2:2 mode, it actually displays text properly (even with sharpness set to 0)! Using this 4:2:2 Standard mode, the 4:2:2/4:4:4 test image shows both 4:2:2 and 4:4:4, but 4:4:4 edges have just a barely noticeable discoloration. Chroma upscaling/doubling, Luma doubling, image upscaling/downscaling are all targeting 4:2:0 - 4:4:4 chroma subsampling conversion. However, I can most definitely see a huge improvement in Standard 4:2:2 mode in video quality and sharpness when using error diffusion, NNEDI3 chroma and luma doubling/upscaling, etc. So, what exactly am I losing?

Just use what your eyes judge to be the best setting.

Originally Posted by XMonarchY

5. I know TotalMedia Theatre does not support madVR, but does PowerDVD 13 support it? Can it be used with madVR to render and playback 3D content?

No.

Originally Posted by Shiandow

What do you think of something like this? This is a modified version of the Floyd-Steinberg weights combined with Madshi's weights.

Combined with how?

Originally Posted by nx6

Better customization of MadVR for third-party developers sounds like a good thing to me.

Not sure what you mean. It's already dead easy for 3rd party developers to add support for custom pixel shaders with madVR. Really, it couldn't be any easier. All they have to do is to feed madVR with the custom shader source code. madVR already does all the rest. No other renderer does that.

Originally Posted by turbojet

The original full file shows this in mediainfo:
Frame rate mode : Variable
Original frame rate : 23.976 fps

The cut version that's linked shows:
Frame rate mode : Constant
Frame rate : 29.970 fps
Original frame rate : 23.976 fps

Should madvr be detecting this as 24 fps like it is currently?

madVR doesn't "detect" anything. The fps information is coming from upstream (splitter/decoder).

-------

For anybody interested in which Error Diffusion algorithm madVR should use, here are a couple of test builds:

http://madshi.net/madVRedtest.rar

There are 8 different builds included, based on ideas and suggestions from different users. I've intentionally *not* named them properly, so you can do a real blind test. Let me know which build number(s) you prefer. And ideally, try them in real time first, without using contrast/brightness boosting. This is for your own good, because that's what counts most. Feel free to double check afterwards with screenshots and contrast/brightness boosts.

Will be interesting to hear the results. I am considering offering two different error diffusion variants: One with very low noise but with worm artifacts. And one without worm artifacts but slightly higher noise. So feel free to nominate either one or two builds. If you nominate two builds, one should be low-noise with worm artifacts, and one should be slightly higher noise without worm artifacts.

All 8 test builds operate on 16x16 pixel blocks, using serpentine scanning. The algorithms used are the following, but in a different order than listed here:

(1) original test build 3 - madshi weights, no randomness
(2) same as (1), but with improved edge error spreading
(3) same as (2), plus adding noise before rounding
(4) original test build 5 - random weights
(5) Floyd-Steinberg, no randomness
(6) same as (5), plus adding noise before rounding
(7) Filter Lite with 0.97 weight sum, no randomness
(8) same as (7), plus adding noise before rounding

I'm really excited about hearing your preferences! I know that I have 2 clear favorites.

[huhn]

Originally Posted by Qaq

Forcing "video"? You had to force "Movie" to let madVR IVTC it 29>23.

the file is not telecine and a normal decimate shouldn't work too is repeats a frame every 5 frames it is definitive not 23p

madvr ivtc detracts it as 4:2:2:2:2 so 12i to 5p so 25 fps looks like 25 fps changed to 30 and with a playback info of 23 hz

madvr normally can't handle 4:2:2:2, but how to handle 4:2:2:2:2 ...

[bacondither]

Nevermind...

[Gagorian]

Originally Posted by bacondither

@madshi

Are you sure you haven't mixed up nr 7 and 8.

Number 7 look like it has added random thresholds and 8 not.

Originally Posted by madshi

I've intentionally *not* named them properly, so you can do a real blind test.
....
The algorithms used are the following, but in a different order than listed here:

Read above?

[Shiandow]

Originally Posted by madshi

Combined with how?

Sorry for being unclear, I was still testing how exactly to combine the two. The image I posted was generated by using the following weights 2/3 of the time:

((0,1/16),
(0,3/16),
(0,5/16),
(7/16,0))

And the following weights the other 1/3 of the time:

((0,0),
(0,0.0090),
(0,0.4861),
(0.7770,-0.3098))

For technical reasons these array are transposed, the rows correspond to the x direction and the columns correspond to the y direction. I also alternated scanning the image left to right and right to left.

Since then I've been testing it in various situations with various different levels of 'mixing'. And I think that this method might be superior even to just using one of both set of weights. It combines the good parts of Flloyd-Steinberg with the good parts of the weights from Build 3, and it adds at least some randomness which can avoid some unwanted patterns.

In the following image I use this method, where I used Build 3 weights 1/5 of the time, on the top part and I use the weights from Build 3 on the lower part:

comparison

The combined method seems to avoid some of the obvious horizontal bands. IMHO this makes the combined method seem smoother, while adding a minimal amount of noise.

[Mano]

Hi, for some files there are 2 audio stream

Audio: FLAC 48000Hz stereo 2304kbps [Japanese, Main - 2.0 FLAC (Audio 1) [Default]]
Audio: AAC 48000Hz stereo [Japanese, Commentary - 2.0 AAC (Audio 2)]

but i cannot switch the audio at all since the play>audio option is grayed.

[bacondither]

After som Quick and dirty enhanced black clipping patterns test(will do a longer test some time later).


Best to worst in the "noisy" artifact free ED category:

1st place(best). Shared by build 7 and build 5.

2nd place. Build 3.

3rd place(worst). Build 1. Ugly dots!


Best to worst in the lower noise and worms are acceptable ED category:

1st place(best). Build 6.

2nd place. Shared by Build 8 and Build 4.

3rd place(worst). Build 2. Ugly dots!

//edit

Build 5 gives me slightly lower rendering times then build 7 and looks identical too. For now i think build 5 is the best choice.

[leeperry]

I didn't read other posts but I think my favorites are 2 and 7, 2 giving the best depth impression as it would appear and 7 looking much less noisy and even better thanks to the lower subjective noise floor but phew...I'll do that again tomorrow morning with fresh eyes if you don't mind.

How about providing those 8 different options so OCD'ed end-users could have a ball and pick whichever one they like best depending on their taste and equipment? It seems only logical that ppl with 6bit TN might not have the same needs as those on real 10bit displays.

[huhn]

Originally Posted by Mano

Hi, for some files there are 2 audio stream


but i cannot switch the audio at all since the play>audio option is grayed.

a audio stream has nothing to do with madvr check your player/splitter.

[n3w813]

Originally Posted by cyberbeing

I'm unsure exactly what you are asking. If you've verified by measurements that your black point is higher when your 3DLUT is active, then it is. If you put your eyes close to a black screen, more pixels should be lit then when you've disabled that 3DLUT. From a distance it should all blend into a single shade, and simply appear brighter in a dark room.

Sorry, I don't think I was clear in my post. I do use a 3DLUT in MadVR. When using the 3DLUT, my black point is raised just a bit from without a 3DLUT. So if I enable error diffusion (DC3 or DC5), I should be able to see the colored 'dots' on a 16-16-16 pattern right? The problem I have is that I don't see ANY dots at all. I'm not trying to disprove what you see but trying to understand if there is something different going on with my set up.

[leeperry]

And anyway, you'll provide a user manual with mVR 1.0 so I don't see how it could be such a big deal to have 8 different ED algorithms if who does what is properly documented IMHO....especially if the ED algorithm is an optional sub-option only meant for power-users(read utterly OCD'ed videophiles ^^).

Better have too many options than not enough, you'll need to explain about Jinc & NNEDI so it can't get much worse and I believe mVR is meant to be the ultimate VR.

Many audio apps come with different dithering algorithms that you can pick and finetune depending on your needs: iZotope MBIT+ Dither

[drew_afx]

Just going by the looks here watching 422 prores hq trailer on 6bit+AFRC e-ips 23" monitor from 1m away

I liked 1, 4, 8

1 seemed to be most ordinary, nothing bad or good

4 showed the most detail and temporal resolution
(but could spot noise on source more easily as well)

8 gave me the pop effect
(but sometimes it seemed less detailed than 4 or 1)

It might not be balanced since I could point out the difference, so I'd say a compromise between 4 and 8 would be the best for me.
Then again I could be very wrong mind tricks me when watching same trailer 8times

[6233638]

Originally Posted by madshi

For anybody interested in which Error Diffusion algorithm madVR should use, here are a couple of test builds:

http://madshi.net/madVRedtest.rar

From a limited amount of testing, looking at noise near black (where I find it most obvious)

1, 3, 5 all show noise on black for me, with 1 & 5 being very noisy.
5 has a tendency to show patterns in the noise.

6 & 7 seem to be very noisy (though not on black) and produce a much darker image near black than the other algorithms. (I've triple-checked this, and get the same results each time)

2 seems to be the best choice for low noise, and avoiding obvious patterns.
2 & 4 produce similarly low-noise images, but 4 seems more prone to introducing obvious patterns.
8 seems to be a bit noisier than 2 or 4.

So #2 seems to be a clear winner for me, at least from the testing I've done so far.

Originally Posted by n3w813

Sorry, I don't think I was clear in my post. I do use a 3DLUT in MadVR. When using the 3DLUT, my black point is raised just a bit from without a 3DLUT. So if I enable error diffusion (DC3 or DC5), I should be able to see the colored 'dots' on a 16-16-16 pattern right? The problem I have is that I don't see ANY dots at all. I'm not trying to disprove what you see but trying to understand if there is something different going on with my set up.

dispcalGUI now has an official workaround to fix your LUT black levels, but you're going to need to create a new 3DLUT for it to work.

[James Freeman]

I can't tell the difference between all of them with my own eyes.
They all look much smoother than the random dither colorful noise, which is very visible.
Funny how the answers from users are all over the place... it may very well be negligible.

Originally Posted by bacondither

After som Quick and dirty enhanced black clipping patterns test

Posted photoshop pictures in no time.
Did you even bother to take the visual test or you simply don't care?

Originally Posted by James Freeman

About 3DLUTs.

I calibrate to Gamma 2.2 Relative and custom white point (around 5900K),
Matrix+Curves (not LUT) which I find gives me a smoother and more correct colors.

Then I create the 3DLUT with BT.1886 2.2 Relative & Luminance Matched Appearance.
Anything else as suggested in the AVS thread resulting in big error in colors and gamma (lifted blacks, blue whites, and more...).

The settings suggested in the AVS thread work only with LUT profiling which I don't use.
I see "color steps" or gradients between colors with LUT, with Matrix its smooth for me.

No dithered blacks with these settings on all DirectComputing versions.

No. Everything is perfectly fine - no raised blacks what-so-ever using 1DLUTs or 3DLUTs. Period. I already explained there is a new feature in dispcalGUI that brings your black level to 0,0,0 RGB coordinates.

[James Freeman]

Originally Posted by XMonarchY

No. Everything is perfectly fine - no raised blacks what-so-ever using 1DLUTs or 3DLUTs. Period. I already explained there is a new feature in dispcalGUI that brings your black level to 0,0,0 RGB coordinates.

Does it also solve the blue whites problem?

[The 8472]

As promised earlier, I'm trying to make some more test patterns

Everything is encoded in 16bit yuv ffv1, i.e. lossless.

r, g, b and white light sources in each corner and a black in the center. stepping up intensity of the corner sources with each frame
1440p
1080p

black-green only gradient:
1440p
1080p

flat tiles with steps of grey smaller than the 8bit quantum (12bit quantum):
1440p
1080p

Best viewed with smooth motion off.