madVR - high quality video renderer (GPU assisted)

[madshi]

Originally Posted by 6233638

I might not get a chance to do much testing until the weekend now, but I'll do what I can.

Can we get a build which lets us customize the keyboard shortcuts though? I find them increasingly difficult to use.

I'd much rather be using QA, WS, ED, RF etc.

Sorry, no easy way to customize the keyboard shortcuts. What is so hard about using them?

I've intentionally released this test build *before* the weekend, so that you guys have it before the weekend starts, so no part of the weekend is lost...

Originally Posted by AngelGraves13

Setting view does not go away. If I enable it, it just floats forever on top of the picture.

That is intentional for the test builds. It stays on for 20 seconds. The purpose is that you have a lot of time to play with the parameters. If it went away after 3-4 seconds, like all other OSD messages, that would make tweaking the parameters more difficult. The OSD message will be cut down to 3-4 seconds when the next official build is released.

[6233638]

Originally Posted by madshi

Sorry, no easy way to customize the keyboard shortcuts. What is so hard about using them?

Well I don't have a number pad or function keys on my keyboard for one thing. So Ctrl+Alt+Shift+F9 means I'm pressing five keys at once. (F9 becomes Fn+9)
I can touch-type, but I have difficulty doing Ctrl+Alt+Shift 1-9 without looking at the keyboard, instead of looking at the screen while making adjustments.

With AltGr instead of Ctrl+Alt, and something like
QWERT: +
ASDFG: −
ZXCVB: on/off

You're only ever pressing two keys, and shouldn't have to look at the keyboard.

[madshi]

Well, there aren't so many letter keys left which are next to each other and not already in use by default madVR settings. Ok, I've modified F9/F10 to K/L now. But I don't have a replacement for the number keys.

http://madshi.net/madVRdeband7.rar

This build also has a bugfix in it, so please everybody who uses the debanding test build update to this one. The keys for the "angleBoost" are now Ctrl+Alt+K/L and Ctrl+Alt+Shift+K/L.

[ryrynz]

Angleboost is amazing.

Here's my offer for the low preset.

6233638, I have tested this vs your previous preference and have found these settings provide better debanding and much improved shadow detail retention.
It could probably still tweak a bit here and there like anything, but overall I'm quite impressed. A excellent find with this setting Madshi!

After some more testing I conclude Localcontrast is not required for great results.

		avgDif	maxDif	grad	penalty  localcontrast   angleboost
low:		0.6	2.4	1.0	1.2	 off             2.0

Originally Posted by madshi

This build also has a bugfix in it

What was fixed? does it affect picture quality?

[madshi]

Originally Posted by ryrynz

Angleboost is amazing.

FWIW, the AngleBoost feature isn't perfect. It misses some parts where debanding should be applied. So we should make sure that we don't lower the other parameters too much. The other parameters should still do useful debanding even without AngleBoost. Maybe your suggested settings already do, haven't checked. Just wanted to say...

Originally Posted by ryrynz

What was fixed? does it affect picture quality?

I had nonGradiantPenalty and localContrast swapped in the OSD. Meaning if the OSD claimed you had localContrast off, you actually had nonGradiantPenalty off and vice versa. I'm sorry about the confusion!

[ryrynz]

Originally Posted by madshi

FWIW, the AngleBoost feature isn't perfect. It misses some parts where debanding should be applied. So we should make sure that we don't lower the other parameters too much. The other parameters should still do useful debanding even without AngleBoost. Maybe your suggested settings already do, haven't checked. Just wanted to say...

Maybe not but it made it jump up to a whole new level, I will continue to tinker with it, I'm hoping madVR can have the low preset enabled by default.

Originally Posted by madshi

I had nonGradiantPenalty and localContrast swapped in the OSD. Meaning if the OSD claimed you had localContrast off, you actually had nonGradiantPenalty off and vice versa. I'm sorry about the confusion!

Ah that explains the lack of difference I saw with localcontrast disabled.. anyway did another test with build 7 and I'm on the fence regarding localcontrast, it does so little I'm not sure it's worth having.
I might enable if it was an available option and I had the GPU power for it, I'm not going to push for the option to be there though, have adjusted my preset accordingly.

Making a comparison vs my old f3kd settings shows my current low preset providing vastly better detail retention whilst maintaining about the same deband strength, awesome stuff.

[Ver Greeneyes]

Originally Posted by madshi

FWIW, the AngleBoost feature isn't perfect. It misses some parts where debanding should be applied. So we should make sure that we don't lower the other parameters too much.

It might not matter much, but have you concidered changing AngleBoost from a binary threshold to a continuum? i.e. use the degree to which the angles match (from -100% to 100% say) to decide how much to boost or penalize the strength of the other settings?

(you could even apply a knee strength setting to the result, like in dynamic range compression, but that's probably getting too fancy for how much good it'll do)

[madshi]

So two suggestions:

(1) penalty vs. boost: I think a penalty for non-matching angles would make things too complicated. Sometimes the angles don't match, but still debanding is needed. So I think a pure boost is a better solution. This way we can set up the other thresholds to do a minimum of debanding in case the angles don't match, and we can rely on that this minimum amount of debanding is always applied.

(2) binary threshold vs continuum: I've actually thought about this, too. All the other thresholds are already applied with "FuzzyLogic" (continuum instead of binary threshold, using a simplified knee strength algorithm). The AngleBoost is currently a binary threshold. I could make the AngleBoost non-binary, too, but I'm not sure if I should do that. FWIW, there are 4 surrounding pixels I have to take into account (each with its own angle), that makes it different to find a continuum formula which covers all the possible situations. But I'll go experiment a little, maybe I can find a solution which works well enough...

[madshi]

Ok, here's the latest build:

http://madshi.net/madVRdeband8.rar

Changes:

(1) The AngleBoost is now of variable strength, depending on how much the angle differs (suggested by Ver Greeneyes).
(2) You can now define both the AngleBoost strength (Ctrl+Alt+H/J), and also the "MaxAngle" for with the boost is applied (Ctrl+Alt+K/L).
(3) You can still turn the AngleBoost completely on/off (Ctrl+Alt+Shift+H/J).
(4) You can activate a debug mode (Ctrl+Alt+Shift+K/L) which shows to which image areas the AngleBoost is applied. White image areas get the full boost. Black image areas don't get any boost. Gray areas get a reduced boost. That should help figuring out the optimal "MaxAngle".

The current default values are "AngleBoost = 2.0" and "MaxAngle = 0.2". Practically this means that if the angle of a gradient differs by up to 13.333%, all thresholds are multiplied by 2x. If the angle differs by 20%, all thresholds are unchanged. If the angle differs by a value between 13.333% and 20% the thresholds are multiplied by a value between 1.0 (20%) and 2.0 (13.333%). For example, an angle difference of 15% would result in a threshold boost of 1.75x.

My first tests indicate that for some gradients the "MaxAngle" could be set as low as 0.01. However, for other gradients, we need at least 0.10-0.15. Maybe even 0.20. Let me know what you think.

[Francois76l]

Hello Madshi!
First, thanks for your amazing work along the years!
I've one question concerning the smooth motion option under madVr v0.86.11.
On my computer (i7 3770k @4.6Ghz + 660GTX Ti + 8Go ram) The picture quality is bad with this option (@60Hz). There is no difference in the picture fluidity and sometimes I can see some frame doubling. It's really strange.

I'm using Windows 7 Pro N (special version without Windows Media Player and Media Center) Do you think my problem can be link?
Thanks again for your help

Originally Posted by madshi

Are you using Windows mode, Overlay mode or Fullscreen Exclusive Mode? For some people, smooth motion only really reliably works in Fullscreen Exclusive Mode. So try whether that fixes the problem. If not, does the OSD report presentation glitches or dropped frames?

Hello Madshi. Thanks for your answer.
I've no dropped frames and no glitches (ctrl+j) and I'm using Fullscreen Exclusive Mode. Sorry for my poor explanations in my first question.

I'm using MPC-BE + LAV Splitter + LAV Video decoder (8 bits YV12 output) + ffdshow with some avisynth scripts (4:4:4 AYUV output) + MadVr 0.86.11

On Madvr:
- Chroma upscaling Jinc 3taps + anti ringing
- Image downscaling Jinc 8taps + anti ringing
- Nothing in "Trade quality for performance"
- Smooth motion On...
- CPU queue size: 12 / GPU: 8

I've also remarked 1 thing. On my Benq W1070 if I output 16 bits colorspace (P016, P216 or Y416) my black are little grayed compared to a 8 bits output. am I crazy?

Thanks again for your explanations Madshi!

[Werewolfy]

Originally Posted by madshi

Ok, here's the latest build:

http://madshi.net/madVRdeband8.rar

Changes:

(1) The AngleBoost is now of variable strength, depending on how much the angle differs (suggested by Ver Greeneyes).
(2) You can now define both the AngleBoost strength (Ctrl+Alt+H/J), and also the "MaxAngle" for with the boost is applied (Ctrl+Alt+K/L).
(3) You can still turn the AngleBoost completely on/off (Ctrl+Alt+Shift+H/J).
(4) You can activate a debug mode (Ctrl+Alt+Shift+K/L) which shows to which image areas the AngleBoost is applied. White image areas get the full boost. Black image areas don't get any boost. Gray areas get a reduced boost. That should help figuring out the optimal "MaxAngle".

The current default values are "AngleBoost = 2.0" and "MaxAngle = 0.2". Practically this means that if the angle of a gradient differs by up to 13.333%, all thresholds are multiplied by 2x. If the angle differs by 20%, all thresholds are unchanged. If the angle differs by a value between 13.333% and 20% the thresholds are multiplied by a value between 1.0 (20%) and 2.0 (13.333%). For example, an angle difference of 15% would result in a threshold boost of 1.75x.

My first tests indicate that for some gradients the "MaxAngle" could be set as low as 0.01. However, for other gradients, we need at least 0.10-0.15. Maybe even 0.20. Let me know what you think.

I did some tests with your latest build and AngleBoost seems to be a great idea! It allows a better debanding with less details blurred. A value of 0.20 for MaxAngle seems to be a good choice, I found one case where 0.15 is not enough and no case where 0.20 is too low.

Here are my settings for low :
avgDif 0.6, maxDif : 2.0, gradient 1.3, nonGradientPenalty : 1.2, localContrast ; off, angleBoost : 2.0, maxAngle : 0.20 .

1.0 for gradient is not enough sometimes like I saw in a earlier post and we can lower maxDif at 2.0. The debanding is better than before with my low settings (0.7, 2.4, 1.5, 1.2, off) and details are better preserved.

Great job madshi

[mandarinka]

This might eventually become interesting for smooth video playback, in the future: http://www.pcper.com/news/Graphics-C...tor-Technology

Basically it makes the rendering on an LCD fully VFR. Sadly, it needs special logic in the monitor, might be nvidia-only feature, and what's worst, will cost extra (so it might end up in highend devices only).

[bacondither]

The new angleBoost algorithm seems to work good some times and bad other times. Especially it seems to blur sharp dark lines alot, and it seems to detect dark grey flat colored boxes as gradients.

It seems to be effective at removing banding in heavy banded material at the expense of blurring sharp dark details.
I would prefer it off on a low deband setting and maby turn it on at medium or high.

I've concocted a test image with some dark boxes of different sizes and a banded gradient:
Test image

Contrast boosted (easier to spot differences) images with different settings on the previously mentioned test image:
No debanding

(My prefered settings "low") avgDif0.7, maxDif 2.4, gradient 1.5, nonGradientPenalty 1.2, localcontrast off, angleBoost off

avgDif0.7, maxDif 2.4, gradient 1.5, nonGradientPenalty 1.2, localcontrast off, angleBoost 2.0, maxAngle 0.20

avgDif0.6, maxDif 2.0, gradient 1.3, nonGradientPenalty 1.2, localcontrast off, angleBoost 2.0, maxAngle 0.20

avgDif0.6, maxDif 2.0, gradient 1.3, nonGradientPenalty 1.2, localcontrast off, angleBoost 2.0, maxAngle 0.01

angleBoost debug mode

/Addendum
I was looking for a medium setting and made a full screen gradient with <7-bit (<127 steps) to simulate bitdeph loss. And i found that the settings below was sufficient to smooth the stairstep gradient to a continuous smooth gradient.

My suggestion for medium debanding:
avgDif 1.1, maxDif 3.0, gradient 2.5, nonGradientPenalty 1.3, localContast off, angleBoost off

[Werewolfy]

I haven't seen this issue... Does it do that even with lower values? 2.0 for angleBoost is maybe too much.

Can you send here your test video?

[iSunrise]

Originally Posted by mandarinka

This might eventually become interesting for smooth video playback, in the future: http://www.pcper.com/news/Graphics-C...tor-Technology

Basically it makes the rendering on an LCD fully VFR. Sadly, it needs special logic in the monitor, might be nvidia-only feature, and what's worst, will cost extra (so it might end up in highend devices only).

Damn you´re fast. Also wanted to post it and ask madshi, what he thinks this could mean regarding madVR and video playback.

Since Nvidia hasn´t yet detailed how exactly G-Sync operates (they said it switches to 3D-Mode) it seems a bit early to tell. However, since madVR basically behaves like a game that uses a D3D-variant and supports fullscreen exklusive, this seems like a no-brainer at first.

Perfectly smooth playback with any input fps at (hopefully) sanely priced monitors (as long as the panel is fast enough) just seems too good to be true.

[e-t172]

Yes, G-SYNC seems very exciting. What's more, that would probably mean that the application has full control over the display clock and is not constrained by the video clock anymore, which would remove audio/video clock drift issues and thus render ReClock completely useless!

[turbojet]

It would be nice if nvidia worked with competitors on a standard. So we don't get 3 different technologies programs need to support which reduces the chances of supporting it. Then of course we'll probably get some things that only support one technology (paid off?). Comparable to cuvid, quicksync and opencl vs dxva.

		avgDif	maxDif	grad	penalty	contrast
low:		0.7	2.4	1.6	1.2	1.4
maximum:	1.9	3.7	4.3	off	off

Maximum is giving me really nice results without any noticeable banding in normal viewing. It's surprising how much improvement it gives to the not so great material (US HDTV) I watch, much more than f3kdb. Good work!

[madshi]

Originally Posted by Francois76l

I've no dropped frames and no glitches (ctrl+j) and I'm using Fullscreen Exclusive Mode.

Have you tried this only on your projector or also on your computer monitor? I'm not sure why SmoothMotion FRC doesn't seem to work as well as you expect on your projector. It could be that something is not working as intended. But maybe also your expectations are too high? SmoothMotion FRC does *NOT* aim to make the source material smoother than it is. All SmoothMotion FRC tries to do is present the original content exactly as it was shot. Native 24p content is not fully smooth by itself. Especially fast horizontal camera pans must judder (unless they're blurred like crazy). And SmoothMotion FRC doesn't even try to improve that.

Please try playing the 24fps Smooth Motion test patterns (see madTestPatternSource download in the first post of this thread) with your projector set to 60Hz. Play that test pattern and then toggle SmoothMotion FRC between on and off during playback, ideally with keyboard shortcuts. Do you see a difference in smoothness that way? It should be smoother with SmoothMotion FRC on. But it will still not be perfectly smooth because 24fps cannot be perfectly smooth.

Originally Posted by Francois76l

I've also remarked 1 thing. On my Benq W1070 if I output 16 bits colorspace (P016, P216 or Y416) my black are little grayed compared to a 8 bits output. am I crazy?

You have a complicated setup with ffdshow and avisynth in between. Try disabling all that extra processing to see if that fixes the gray blacks. If it does, something is wrong with your ffdshow/avisynth/madVR configuration.

Originally Posted by mandarinka

This might eventually become interesting for smooth video playback, in the future: http://www.pcper.com/news/Graphics-C...tor-Technology

Basically it makes the rendering on an LCD fully VFR. Sadly, it needs special logic in the monitor, might be nvidia-only feature, and what's worst, will cost extra (so it might end up in highend devices only).

I'm not sure if this allows changing the refresh rate dynamically, all the time. Or maybe whether it just allows the monitor to support any refresh rate (but not dynamic switching).

Originally Posted by turbojet

It would be nice if nvidia worked with competitors on a standard. So we don't get 3 different technologies programs need to support which reduces the chances of supporting it. Then of course we'll probably get some things that only support one technology (paid off?). Comparable to cuvid, quicksync and opencl vs dxva.

Very true and I fully agree. NVidia already tried to sell their own 3D technology and IMHO it utterly failed. Now I hear that this G-Sync will come with license costs etc. So there will not be an open standard which AMD and Intel get join. No thanks, doesn't sound interesting for me.

What I'd like to see is for the driver to support very small changes to the pixel clock without a full mode reset. That would truely render ReClock useless (unless you want to modify the source fps), and it should work with most displays without any changes. That would be a much better solution for video playback. G-Sync is aimed at gamers, not at HTPC users.

Originally Posted by Werewolfy

I did some tests with your latest build and AngleBoost seems to be a great idea! It allows a better debanding with less details blurred. A value of 0.20 for MaxAngle seems to be a good choice, I found one case where 0.15 is not enough and no case where 0.20 is too low.

Here are my settings for low :
avgDif 0.6, maxDif : 2.0, gradient 1.3, nonGradientPenalty : 1.2, localContrast ; off, angleBoost : 2.0, maxAngle : 0.20 .

1.0 for gradient is not enough sometimes like I saw in a earlier post and we can lower maxDif at 2.0. The debanding is better than before with my low settings (0.7, 2.4, 1.5, 1.2, off) and details are better preserved.

Great job madshi

Glad to hear you like it!

Originally Posted by turbojet

Maximum is giving me really nice results without any noticeable banding in normal viewing. It's surprising how much improvement it gives to the not so great material (US HDTV) I watch, much more than f3kdb. Good work!

Originally Posted by bacondither

The new angleBoost algorithm seems to work good some times and bad other times. Especially it seems to blur sharp dark lines alot, and it seems to detect dark grey flat colored boxes as gradients.

It seems to be effective at removing banding in heavy banded material at the expense of blurring sharp dark details.
I would prefer it off on a low deband setting and maby turn it on at medium or high.

I've concocted a test image with some dark boxes of different sizes and a banded gradient:
Test image

Contrast boosted (easier to spot differences) images with different settings on the previously mentioned test image:
No debanding

(My prefered settings "low") avgDif0.7, maxDif 2.4, gradient 1.5, nonGradientPenalty 1.2, localcontrast off, angleBoost off

avgDif0.7, maxDif 2.4, gradient 1.5, nonGradientPenalty 1.2, localcontrast off, angleBoost 2.0, maxAngle 0.20

avgDif0.6, maxDif 2.0, gradient 1.3, nonGradientPenalty 1.2, localcontrast off, angleBoost 2.0, maxAngle 0.20

avgDif0.6, maxDif 2.0, gradient 1.3, nonGradientPenalty 1.2, localcontrast off, angleBoost 2.0, maxAngle 0.01

angleBoost debug mode

/Addendum
I was looking for a medium setting and made a full screen gradient with <7-bit (<127 steps) to simulate bitdeph loss. And i found that the settings below was sufficient to smooth the stairstep gradient to a continuous smooth gradient.

My suggestion for medium debanding:
avgDif 1.1, maxDif 3.0, gradient 2.5, nonGradientPenalty 1.3, localContast off, angleBoost off

I'm not sure if your test image qualifies for showing up problems with AngleBoost. Why would those dark gray boxes classify as "image detail" and not as "banding"? Do you consider them "image detail" because they are "closed" at the bottom and top, forming them into a geometrical rectangle? What would you say if the same shades of gray would go from the bottom to the top of the screen, resulting in vertical stripes alternating between black and very dark? Would that still be "image detail" or would you consider that "banding"?

From my point of view large flat image areas which are barely different from neighboring large flat image areas are more likely to be "banding" than "image detail". Image detail usually contains a little bit of contrast, a little bit of per pixel changes, a little bit of angles/curves/edges in it. There's nothing of that in your large flat barely visible gray boxes. Because of that I don't think those boxes can be considered "image detail".

Can you provide a real life sample (or 2 or 3) which shows the following problems:

> Especially it seems to blur sharp dark lines alot,
> and it seems to detect dark grey flat colored
> boxes as gradients.
> It seems to be effective at removing banding
> in heavy banded material at the expense of
> blurring sharp dark details.

Having a sample (or 2 or 3) which show these problems would be really useful. Maybe I can find a way to improve the AngleBoost feature accordingly.

There's one thing making me wonder, though: In your "debug" mode image it seems that only the left border of your gray boxes is detected at all, but not the right border. That's weird, will look into that. Doing so will not change the final output, though.

[iSunrise]

Originally Posted by madshi

I'm not sure if this allows changing the refresh rate dynamically, all the time. Or maybe whether it just allows the monitor to support any refresh rate (but not dynamic switching).

According to the demos and setups they have used at the event, you can do dynamic changes as well as keep a steady and constant refresh locked.

Here´s a video of it:
http://www.youtube.com/watch?v=WiZ9l3I26ec

The one clear advantage I see is that you could drive G-SYNC equipped monitors at native movie refresh rates (where you would have to buy much more expensive monitors like Eizo CGs) or even some movie refresh rate multiples that are only available with the G-Sync module. That´s because the module completely replaces the scaler component inside the monitor and the GPU and pixel clock drives it. And depending how well they implemented it, this could also lead to other side-benefits like the ones you mentioned (smaller variations in pixel clocks, since they will possibly do a bit of a rewrite of that part of the driver). Maybe there are even more benefits we don´t even know about, more details will probably be available on the 21st and if it works like they made it sound, it seemed like you wouldn´t even need to adjust anything within madVR, since it behaves like a game, but who knows at this point.

I agree though that an open standard would have been nice, too.

BTW, love the new deband, even though I don´t watch a lot of Anime and I am certainly not the prime target for this.

[6233638]

Originally Posted by madshi

I'm not sure if this allows changing the refresh rate dynamically, all the time. Or maybe whether it just allows the monitor to support any refresh rate (but not dynamic switching).

As I understood it, G-Sync means that the display's refresh rate is controlled by the graphics card - when a new frame is drawn, the display updates.

This allows you to avoid screen tearing without increasing latency or limiting framerates. (e.g. locking to 30 or 60fps)

EDIT: Yes, it's continuously variable and synced to the framerate: http://www.youtube.com/watch?v=WiZ9l3I26ec
That said, I'm not sure how uneven framerates, even with the refresh rate synced, and without tearing, are going to avoid judder.

Originally Posted by madshi

Very true and I fully agree. NVidia already tried to sell their own 3D technology and IMHO it utterly failed. Now I hear that this G-Sync will come with license costs etc. So there will not be an open standard which AMD and Intel get join. No thanks, doesn't sound interesting for me.

LightBoost monitors seem to be in high demand from PC gamers - they are significantly better than any other display on the market when it comes to latency and motion handling. The only other comparable displays are CRTs.