madVR - high quality video renderer (GPU assisted)

[THX-UltraII]

I m not talking about the problem FULL and LIMITED. This already works out-of-the-box with the latest Intel driver (you can simply select AUTO, FULL or LIMITED in the Intel Control Panel and it works perfect).

I am talking about huhn saying that Intel cards output ycbcr and not rgb and that this is supposed to affect PQ in a bad way when using madVR.

[andybkma]

Originally Posted by THX-UltraII

I m not talking about the problem FULL and LIMITED. This already works out-of-the-box with the latest Intel driver (you can simply select AUTO, FULL or LIMITED in the Intel Control Panel and it works perfect).

I don't have these options in my Intel driver control panel. Where exactly are they located? I am already using the latest Intel drivers for my HD4000 with Win7 64bit that I could find on Intel's site (9/11/2013, 15.31.17.64.3257). I don't have any auto, full or limited options or I need clear cut directions to find them ;-)

Also, my post wasn't directed at you but mVR users in general who have struggled as I have to achieve proper blacks with their HTPC....

[Werewolfy]

Originally Posted by 6233638

If you don't want any detail loss, then you have to disable debanding. It's all a balance between detail loss and banding.

The "wipers" themselves are just big macroblocks in your sample though, all one solid color. Here it is brightened up:



I wouldn't call it detail loss, just debanding at work, smoothing the edges.
You might also want to check your display's calibration if those are easy to see without being brightened like that.

You're right, some "details" will always be lost with debanding. I was trying to find a setting without any detail loss but in that case, the debanding strenght is very limited.

Here my current settings :
low : avgDif: 0.7 maxDif: 2.5 gradient: 1.5 nonGradientPenalty : 1.2
mid : avgDif: 1.2 maxDif: 3.2 gradient: 2.8 nonGradientPenalty : 1.3
high : avgDif: 2.3 maxDif: 4.0 gradient: 4.3 nonGradientPenalty : 1.3

For low, the debanding is effective and there is almost no detail loss. At higher values, I find the picture too smooth on certain places. Lower values exhibit a lot more banding.
For mid, the debanding is very effective on most videos. There is some detail loss but for Anime it's a perfect setting. There is less details generally in Anime than movies.
For high, I aim a very effective debanding on bad sources. But I need more bad videos to test it. If anyone has links to that kind of videos, I'd appreciate it.

I like the new nonGradientPenalty setting. It allows to maintains some details that are usually blurred by the gradient setting and maintaining good debanding at low values. But some people might desactive it for high because it maintains sometimes compression artifacts.
I still find the LocalContrast quite useless. There is almost no difference with or without and when there is a difference, it's very minimal.

[madshi]

Interesting, you guys are getting nearer to each other in your parameter choices. @6233638, what is your opinion about Werewolfy's current settings? They're relatively near to yours, but not totally identical. I would really like to get rid of localContrast, but I'm not sure if I can. I do wonder whether I should have 3 or 4 settings. It seems that Werewolfy prefers "avgDif 0.7" and 6233638 prefers "avgDif 0.8" for the "low" setting. Maybe a "very low" setting might make sense. I'm thinking that if we get native 4K content sooner or later with native 10bit decoding then maybe a "very low" setting might make sense to improve 10bit gradients slightly. For clean native 10bit sources "low" might be overkill. Of course we don't have clean native 10bit sources at the moment (do we?), so maybe it's hard to find good parameter values for that...

[bacondither]

I agree on avgDif 0.7 on a low setting, but values >1.4-1.5 for maxDif and >1.3-1.4 for gradient seems to have quite a steep diminishing return on debanding. Atleast for a low setting.

/errata
For some low resolution material it seems that maxDif would have to be atleast 1.8 to have sufficient effect.

With nonGradientPenalty over 1.1 i can see some miniscule blocking and slight banding sometimes. 1.2 would be max. Atleast at a low setting. I would use 1.1 or disabled.

Test scene 1:
No debanding
avgDif 0.7 maxDif 1.5 gradient 1.4
avgDif 0.7 maxDif 2.5 gradient 1.5
avgDif 0.7 maxDif 2.5 gradient 1.5 nonGradientPenalty 1.2

Test scene 2:
No debanding
avgDif 0.7 maxDif 1.5 gradient 1.4
avgDif 0.7 maxDif 2.5 gradient 1.5
avgDif 0.7 maxDif 2.5 gradient 1.5 nonGradientPenalty 1.2

[madshi]

@bacondither,

I think the other guys have also started with relatively low "maxDif" values, but have increased them recently. Of course a lot depends on the source content you test with. With some material low values work ok, with other material higher values are needed. Basically the flatter a banded gradient in the image is, the lower "maxDif" can be. If there is a rather strong gradient in the image, "maxDif" needs to be bigger for debanding to work correctly.

About "nonGradientPenalty": If you see too much blocking/banding with 1.2 or 1.3, then you can either turn "nonGradientPenalty" back to 1.1 or even 1.0. Or alternatively you could leave it at 1.2 or 1.3 and instead turn "gradient" further up, since "gradient" and "nonGradientPenalty" are related. Blocking/banding caused by "nonGradientPenalty" can be removed by increasing "gradient". The big question is which combination of "gradient" and "nonGradientPenalty" produces the better overall result (best debanding with the lowest detail loss). I'm not sure about that.

[Werewolfy]

maxDif doesn't seem to harm the picture so I don't see why reducing it. Of course, maybe it harms the picture but not on the samples I have so feel free to test it on your samples and tell if it harms the pictures with the values I gave for example.

nonGradientPenalty seems to help at low values, I'd have not so high values for Gradient otherwise. The only side effect is that the compression artifacts are not so well hidden when it's activated so maybe it's not suited for the high setting.

Originally Posted by madshi

Interesting, you guys are getting nearer to each other in your parameter choices. @6233638, what is your opinion about Werewolfy's current settings? They're relatively near to yours, but not totally identical. I would really like to get rid of localContrast, but I'm not sure if I can. I do wonder whether I should have 3 or 4 settings. It seems that Werewolfy prefers "avgDif 0.7" and 6233638 prefers "avgDif 0.8" for the "low" setting. Maybe a "very low" setting might make sense. I'm thinking that if we get native 4K content sooner or later with native 10bit decoding then maybe a "very low" setting might make sense to improve 10bit gradients slightly. For clean native 10bit sources "low" might be overkill. Of course we don't have clean native 10bit sources at the moment (do we?), so maybe it's hard to find good parameter values for that...

I thought also you could add a very low setting but now it seems useless to me. It's true that we can find a lower setting than the low setting I gave for example but in that case the debanding effect is quite minmal so yes maybe for native 4K content 10bit sources it can ben useful but for now, I don't really see the point.

[Skankee]

I compared Werewolfy's and 6233638's low-debanding settings on high-quality videos to see if (low) debanding has a negative influence on picture-quality.


There was not much banding in my videos, so i couldn't see a difference between Werewolfy's and 6233638's low-debanding settings. In almost all examples the picture-quality was as good as without debanding, + noticeable debanding in some cases.

But in really dark areas i noticed something i dont like.

For me it looks like the "bright" areas stay bright and the dark areas become much brighter, so that the video seems partly brighter in some areas with active debanding.
I expected that the "bright" areas would be a little bit darker, so that the average luminosity stay the same (visual).

No debanding (part):
http://s1.directupload.net/images/user/131015/6p8eeupz.png

low-debanding (part):
http://s1.directupload.net/images/user/131015/74cb7nwh.png



I looked at the luminosity-histogram and the average luminosity is almost identical. But especially the lower part in the middle looks too bright for me with debanding.

Full-size:
No debanding:
http://s1.directupload.net/images/user/131015/t6uej6ba.png

Werewolfy low-debanding:
http://s14.directupload.net/images/user/131015/v45eysk9.png

6233638 low-debanding:
http://s14.directupload.net/images/user/131015/cc3qxczb.png

[6233638]

Originally Posted by madshi

@6233638, what is your opinion about Werewolfy's current settings? They're relatively near to yours, but not totally identical. I would really like to get rid of localContrast, but I'm not sure if I can.

Unfortunately, I'm still finding sources where localContrast seems to be useful, and it doesn't seem like nonGradientPenalty will completely replace it.

I must say, I'm finding it difficult to see anything but very subtle changes when adjusting nonGradientPenalty with the sources I've looked at, but 1.2 does seem to have some benefit at low settings.
I'm not sure that I would want to use high levels of gradient and nonGradientPenalty together, because high levels of gradient seem to add a lot of dither/noise to the image.

I really need to keep better track of my testing, because I'm not sure what video(s) it was that caused me to raise the low avgDif to 0.8 before
So unless I can find another example of that, 0.7 seems acceptable.
Using 0.7 as the avgDif, 2.4 seems like the logical choice for maxDif, and a gradient of 1.6 seemed to be as low as I could go with those before debanding effectiveness was reduced.

I'm somewhat flexible on the nonGradientPenalty and localContrast settings, but 1.2/1.4 (possibly 1.5) seemed to have little impact on debanding effectiveness, while helping preserve some shadow detail.

		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

I need to do more testing to find the optimal settings for a "medium" preset, but I'm thinking something around an avgDif of 1.2 seems like a good choice. I would not want any lower.

I'm still quite happy with the max settings I posted before, and haven't found a source where Werewolfy's stronger settings are required.

I don't know that you would need separate presets which use localContrast, just have presets with localContrast values (and this would likely only be low/medium) and a "trade quality for performance" that disables the localContrast portion if required.

Originally Posted by bacondither

I agree on avgDif 0.7 on a low setting, but values >1.4-1.5 for maxDif and >1.3-1.4 for gradient seems to have quite a steep diminishing return on debanding. Atleast for a low setting.

I think it depends on what you consider a "low" setting to be useful for.
If we are going to use an avgDif of 0.7, I think maxDif should be 2.4. Going any higher does not result in much improvement for debanding, and going lower is much less effective.

Originally Posted by bacondither

Test scene 1:
Test scene 2:

Can you please post video samples for these. I do not find your example images to have sufficient debanding, and would really like to experiment and see what settings would be effective with them.

Originally Posted by Skankee

I compared Werewolfy's and 6233638's low-debanding settings on high-quality videos to see if (low) debanding has a negative influence on picture-quality

Can you post a video sample of this? It seems like a good test for medium settings.

[bacondither]

Originally Posted by 6233638

Unfortunately, I'm still finding sources where localContrast seems to be useful, and it doesn't seem like nonGradientPenalty will completely replace it.

Could you provide an example of that please?

Originally Posted by 6233638

If we are going to use an avgDif of 0.7, I think maxDif should be 2.4. Going any higher does not result in much improvement for debanding, and going lower is much less effective.

Are those images from the AVS HD 709 grayscale ramp?, and if so i too found maxDif 2.4 to remove most bands in that syntetic test image. But you almost have to be pixel peeping to see it.
It seems that the detail reduction by using maxDif 2.4 over 1.5-1.6 is so miniscule it might be a good vaule for a low setting.

Originally Posted by 6233638

Can you please post video samples for these. I do not find your example images to have sufficient debanding, and would really like to experiment and see what settings would be effective with them.

You cannot eliminate all banding all the time without destroying the image, especially if you try to tune to remove every trace of banding in clips of quite high banding that i will post below.
In a low setting it "should" try to hide the banding from "eeeew banding" to "this looks acceptable", or atleast that is what i think.

The clips:
Test scene 1
Test scene 2

Additional clips:
Test scene 3
Test scene 4

[ryrynz]

Originally Posted by 6233638

Unfortunately, I'm still finding sources where localContrast seems to be useful, and it doesn't seem like nonGradientPenalty will completely replace it.

Agreed, even with localContrast fixed I'd still prefer it to stay, I know it's minor but I'm still seeing a difference.

Originally Posted by 6233638

Using 0.7 as the avgDif, 2.4 seems like the logical choice for maxDif, and a gradient of 1.6 seemed to be as low as I could go with those before debanding effectiveness was reduced.

I've had a play with the low setting using 6233638's parameters as a basis and I've found the following to provide a smoother deband transition. I think we're basically there as far as the low preset goes, please provide feedback.

		avgDif	maxDif	grad	penalty	contrast
low:		0.7	2.4	2.5	1.5	1.5

nonGradientPenalty if it's costly could be removed, the differences it makes are terribly minor (at these settings) and are hard to identify without some detailed image comparison. I do prefer it enabled but I could live without it.

[bacondither]

I think local contrast could be made optional or removed. Makes almost no difference at removing banding at the cost of ~20% longer rendering times.
nonGradientPenalty has seems to have very low impact on rendering times and does improve detail.

This is my reevaluated settings for low.

		avgDif	maxDif	grad	penalty	localcontrast
low:		0.7	2.4	1.5	1.2	off

[madshi]

LOW SETTINGS FINDING:

Originally Posted by Werewolfy

maxDif doesn't seem to harm the picture so I don't see why reducing it. Of course, maybe it harms the picture but not on the samples I have so feel free to test it on your samples and tell if it harms the pictures with the values I gave for example.

Originally Posted by bacondither

Are those images from the AVS HD 709 grayscale ramp?, and if so i too found maxDif 2.4 to remove most bands in that syntetic test image. But you almost have to be pixel peeping to see it.
It seems that the detail reduction by using maxDif 2.4 over 1.5-1.6 is so miniscule it might be a good vaule for a low setting.

Originally Posted by bacondither

nonGradientPenalty has seems to have very low impact on rendering times and does improve detail.

This is my reevaluated settings for low.

		avgDif	maxDif	grad	penalty	localcontrast
low:		0.7	2.4	1.5	1.2	off

Originally Posted by 6233638

I really need to keep better track of my testing, because I'm not sure what video(s) it was that caused me to raise the low avgDif to 0.8 before
So unless I can find another example of that, 0.7 seems acceptable.
Using 0.7 as the avgDif, 2.4 seems like the logical choice for maxDif, and a gradient of 1.6 seemed to be as low as I could go with those before debanding effectiveness was reduced.

Originally Posted by ryrynz

I've had a play with the low setting using 6233638's parameters as a basis and I've found the following to provide a smoother deband transition. I think we're basically there as far as the low preset goes, please provide feedback.

		avgDif	maxDif	grad	penalty	contrast
low:		0.7	2.4	2.5	1.5	1.5

Ok, good, so it seems everybody agrees with "avgDif 0.7" and "maxDif 2.4" for the low setting? That's progress! So let's consider those two fixed. For the low settings we still need to decide on 2 things:

(1) How to set "gradient" and "nonGradientPenalty". Remember: A higher "nonGradientPenalty" requires a higher "gradient".

gradient / nonGradientPenalty:

Werewolfy: 1.5 / 1.2
bacondither: 1.5 / 1.2
6233638: 1.6 / 1.2
ryrynz: 2.5 / 1.5

ryrynz' suggestions are higher than the rest, but he has *both* values raised, so it makes sense, too. Can we somehow get an agreement on these values, too?

(2) We also still need to decide on "localContrast". See next section.

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

LOCAL CONTRAST:

Originally Posted by Werewolfy

I still find the LocalContrast quite useless. There is almost no difference with or without and when there is a difference, it's very minimal.

Originally Posted by bacondither

I think local contrast could be made optional or removed. Makes almost no difference at removing banding at the cost of ~20% longer rendering times.

Originally Posted by 6233638

Unfortunately, I'm still finding sources where localContrast seems to be useful, and it doesn't seem like nonGradientPenalty will completely replace it.

I don't know that you would need separate presets which use localContrast, just have presets with localContrast values (and this would likely only be low/medium) and a "trade quality for performance" that disables the localContrast portion if required.

Originally Posted by ryrynz

Agreed, even with localContrast fixed I'd still prefer it to stay, I know it's minor but I'm still seeing a difference.

I could live with a "trade quality for performance" setting for "localContrast". But could you guys, 6233638 and/or ryrynz, please provide a small test sample which shows the benefit of "localContrast"? Please let us know at which runtime in the sample a difference is visible. Thanks. Maybe that would convince Werewolfy, bacondither and me of its use. And that would also help finding optimal values for "localContrast". It seems that you guys favor a setting of 1.4 or 1.5 for "low"?

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

RANDOM COMMENTS:

Originally Posted by 6233638

I must say, I'm finding it difficult to see anything but very subtle changes when adjusting nonGradientPenalty with the sources I've looked at, but 1.2 does seem to have some benefit at low settings.
I'm not sure that I would want to use high levels of gradient and nonGradientPenalty together, because high levels of gradient seem to add a lot of dither/noise to the image.

That sounds weird to me. The amount of dithering/noise should be totally independent of any of the debanding parameters. The dithering strength used is always the same. If large image areas are exactly between two 8bit values (e.g. 80.5), then the dithering noise might be more visible. But it's really quite random which debanding values put which specific areas to which 8bit values.

Originally Posted by Werewolfy

I thought also you could add a very low setting but now it seems useless to me. It's true that we can find a lower setting than the low setting I gave for example but in that case the debanding effect is quite minmal so yes maybe for native 4K content 10bit sources it can ben useful but for now, I don't really see the point.

Yeah, maybe we should forget about "very low" for now. If you guys are all happy with whatever settings we find for "low", and agree on that detail isn't harmed (in any way worth mentioning) by those settings, then "very low" can wait until we see a need for it.

Originally Posted by 6233638

I'm still quite happy with the max settings I posted before, and haven't found a source where Werewolfy's stronger settings are required.

Let's first decide on the final "low" settings, then we can concentrate more on "mid" and high".

Originally Posted by ryrynz

nonGradientPenalty if it's costly could be removed, the differences it makes are terribly minor (at these settings) and are hard to identify without some detailed image comparison. I do prefer it enabled but I could live without it.

The cost is pretty small.

Originally Posted by Skankee

But in really dark areas i noticed something i dont like.

For me it looks like the "bright" areas stay bright and the dark areas become much brighter, so that the video seems partly brighter in some areas with active debanding.
I expected that the "bright" areas would be a little bit darker, so that the average luminosity stay the same (visual).

No debanding (part):
http://s1.directupload.net/images/user/131015/6p8eeupz.png

low-debanding (part):
http://s1.directupload.net/images/user/131015/74cb7nwh.png

I looked at the luminosity-histogram and the average luminosity is almost identical. But especially the lower part in the middle looks too bright for me with debanding.

Interesting. A key part of the debanding algorithm is to build an average of 4 surrounding pixels. This is one of those cases where linear light might help. I thought that the value difference between those pixels would be small enough to make linear light superfluous, but maybe I was wrong. So here's a new test build which has an additional "linear light" switch. Keyboard shortcuts Ctrl+Alt+Shift+F9/F10:

http://madshi.net/madVRdeband5.rar

On a quick check it doesn't seem to make any difference in most situations, but maybe it helps in your specific case to avoid the brightening effect you mentioned?

[ryrynz]

Double post.

[ryrynz]

Originally Posted by bacondither

I think local contrast could be made optional or removed. Makes almost no difference at removing banding at the cost of ~20% longer rendering times.
nonGradientPenalty has seems to have very low impact on rendering times and does improve detail.

This is my reevaluated settings for low.

		avgDif	maxDif	grad	penalty	localcontrast
low:		0.7	2.4	1.5	1.2	off

Optional yes, removal I most certainly hope not. It makes more of a difference than changing many of the settings we have at our disposal.. I'll happily provide screenshots to show you, it's all in the blending transition. Believe me if I could generate the same quality results with it disabled I'd push for it's removal.

localcontrast makes more of a difference than nonGradientPenalty, again I can happily back this up.

[madshi]

@ryrynz, a short video sample would be great!

[bacondither]

Originally Posted by madshi

ryrynz' suggestions are higher than the rest, but he has *both* values raised, so it makes sense, too. Can we somehow get an agreement on these values, too?

Well, we could take the median of our collective findings.

Variable            N  Median
gradient            4  1,550
nongradientpenalty  4  1,2000

[ryrynz]

It's going to be tough to push such an minor change in image quality with localContrast considering it's impact but the difference is in the softness of the image.
It find the differences easier to see when switching from no deband to LocalContrast and then to no LocalContrast, brightness and contrast has been increased to help highlight the differences.

The fix you implemented Madshi diminished the difference somewhat, I've used screenshots from the video I linked you earlier.

It's this slight softening of detail that would be removing the shadow detail that 6233638 talks about, it's really a case of how pedantic you/we want to be about this light detail.
Since I'm examining things at the pixel level.. it's hard for it not to do unnoticed and I prefer it being there.. by default? maybe not.

No deband


My settings with LocalContrast=1.5


My settings with LocalContrast=disabled

Originally Posted by bacondither

Well, we could take the median of our collective findings.

I'd rather compare and choose based on content, I'm striving for the absolute best deband quality for this because let's face it we don't have a lot of strength to work with and ideally I think we'd want low on by default.
I'm relying on LocalContrast here to maximize the amount of detail we get out of low whilst still maintaining a reasonable blending.

We've already tightened the two largest screws, so just about anything now is good enough really, but I've checked out Nongradientpenalty 1.2 and I'm happy with that so I'm in with everyone else.
I still think the gradient needs to be higher than 1.5-1.6, I've settled at 2.0 for this ATM. LocalContrast at 1.5 is almost identical to 1.4 but 1.5 from what I see has the every so slight edge, 6233638 could you also confirm this for LocalContrast?

[Werewolfy]

I don't really see any difference between Local Contrast on/off in your screenshots. But if you and 6233638 see a difference, there is cetainly one. But it's very minor then and there is a cost so I wouldn't like to have this option by defaut but maybe a "trade quality for performance" option would be a good idea.

I compared 1.5/1.2 vs 1.6/1.2 vs 2.5/1.5 for gradients and nonPenaltyGradient. It's almost the same results. 1.6/1.2 smooths a little bit more some details. 1.5/1.2 and 2.5/1.5 are pretty much the same, I can't figure out which one is the best between those two settings.

Originally Posted by 6233638

I'm still quite happy with the max settings I posted before, and haven't found a source where Werewolfy's stronger settings are required.

I did find a video when there is some benefits especially when maxDif is higher than 3.7 but we'll discuss about that later.

[Skankee]

Originally Posted by 6233638

Can you post a video sample of this? It seems like a good test for medium settings.

~20 MB:
http://www.sendspace.com/file/vxt36d

Originally Posted by madshi

On a quick check it doesn't seem to make any difference in most situations, but maybe it helps in your specific case to avoid the brightening effect you mentioned?

Unfortunately not. I tested linear light with low and high debanding, no difference visible to me.

But i found out that i also get these brighter areas when i use dithering (it is disabled on my previous screenshots because of better blacks and a reduced filesize).

So maybe my problem is related to dither and not to debanding.

With dithering (and disabled debanding) there is noticeable less banding in my video-sample, so maybe it is not so good for medium dithering-tests anymore, i don't know.

But there are still 2 questions left:
-can i disable dithering when i use low debanding (because both do some debanding),
-are my "dark areas" in my test-images a bug because of missing dithering or is dithering doing something "strange" and create too bright areas?