madVR - high quality video renderer (GPU assisted)

[StinDaWg]

I don't know what the heck happened but I'm getting ridiculous rendering times using NNEDI3 after updating to the latest version of madVR. 20ms higher render times on my AMD 7850.

I'm going to restart my computer to see if something got messed up.

Edit: Nope. I'm at a loss for words. I haven't changed anything on my end except to continue to update Windows 8.1.

[flashmozzg]

I don't know what the heck happened but I'm getting ridiculous rendering times using NNEDI3 after updating to the latest version of madVR. 20ms higher render times on my AMD 7850.

I'm going to restart my computer to see if something got messed up.

Edit: Nope. I'm at a loss for words. I haven't changed anything on my end except to continue to update Windows 8.1.

Have you updated madVR itself? Is interop hack enabled?
Have you not installed anything into your PC PCIe slot? NNEDI3 on amd is very PCIe speed dependent. (You can try to see the pcie speed for your gpu in gpu-z)

[Anime Viewer]

I don't know what the heck happened but I'm getting ridiculous rendering times using NNEDI3 after updating to the latest version of madVR. 20ms higher render times on my AMD 7850.

I'm going to restart my computer to see if something got messed up.

Edit: Nope. I'm at a loss for words. I haven't changed anything on my end except to continue to update Windows 8.1.

My system is different from yours in quite a few ways, but I haven't seen any drastic increase between versions.

Did you uninstall the previous version before you installed the new version? If you did uninstall are you sure you re-added all the settings? (Like display resolution and refresh rates?) You used to have have display modes set, and as you know they can be a factor. (http://forum.doom9.org/showthread.php?p=1702999#post1702999) You didn't put any checks in the new features that you might not have used previously (up-scaling refinement: SuperRes, FineSharp, or image enhancement) did you? You didn't uncheck any trade quality for performance check marks you previously had checked, did you?

[StinDaWg]

I rolled back to an old version of madVR and still have high render times so that's not the issue. This started a few days ago after Windows did a system update and rebooted. GPU usage is very high now. 91% with the same NNEDI3 settings where it was around 50% before.

[Anime Viewer]

I rolled back to an old version of madVR and still have high render times so that's not the issue. This started a few days ago after Windows did a system update and rebooted. GPU usage is very high now. 91% with the same NNEDI3 settings where it was around 50% before.

Perhaps it is NNEDI3 related since I haven't used NNEDI3 in quite a while, and I didn't have that problem (I've been using NEDI). Are you using NNEDI3 for only image doubling, or something else? (I know NNEDI3 for Chroma upscaling can increase render times x4.

[StinDaWg]

Only doubling.

[Werewolfy]

@madshi

I tested the latest version of Shiandow's debanding and I still prefer the high preset. Most of the time, there is no clear winner but I found an example where the high preset is clearly superior, you have to set very high values in Shiandow's debanding to be on par with madVR's debanding and by doing that you lose a lot of details.

High preset : [img=http://s5.postimg.org/6enu7he9z/high1.png] (http://postimg.org/image/q99vtlthf/full/)
[/url]
Shiandow - power : 0.80 - margin : 0.00 : [img=http://s5.postimg.org/qnb7t7dl3/shiandow085000.png] (http://postimage.org/index.php?lang=french)

Shiandow - power : 0.85 - margin 0.30 : [img=http://s5.postimg.org/9ckt16lxj/shiandow085030.png] (http://postimage.org/index.php?lang=french)

Shiandow - power : 0.90 - margin : 0.50 : [img=http://s5.postimg.org/ushj1dlzb/shiandow090040.png] (http://postimage.org/index.php?lang=french)



Sometimes Shiandow's algo is better, sometimes it's weaker. MadVR's debanding smoothes the banding and Shiandow's algo blurs it. It also blurs compression artefacts which could be usefull but it's not really the role of a debanding algo and I think that a denoise or a deblocker will be a lot more efficient at this task.

I also played with the high preset and I found a way to improve it a little bit by setting Angleboost at 1.4 and maxAngle at 0.10.
You don't lose more details than with the original high preset and it can help sometimes.

[img=http://s5.postimg.org/gnbpzkcxz/highc1.png] (http://postimage.org/index.php?lang=french)
[url=http://postimage.org/index.php?lang=french] (don't pay attention to avgDif, it's a mistake and it dosen't change anything)

I know that the difference is not huge but you don't lose more details by doing that so I don't really see why not doing it ;) With these settings, I find hard to find an example where Shiandow's algo is better.

[InstantAli3n]

I must agree with Werewolfy. After switching between the algo's for a few days I've come to the conclusion that madVR's high preset does a better job with less blurring.

n00b question: Is it normal to get frame drops above about 85% GPU usage? I don't reach 90% at all but if I use any options that get me between 82-87% I get frame drops. :(

GTX 770, 347.88

[Asmodian]

I have wanted to provide feedback on debanding but my experiments have failed to provide useful distinctions (unless a report of inconclusive is useful). :o

Depending on the source and the exact banding and details I am paying attention to I can sometimes notice differences but I have not found anything that would allow me to feel comfortable saying generally one is better than the other.

I generally use no or light debanding, conceptually I prefer viewing banding to losing details. I haven't done any blind tests to determine what I actually prefer. :devil:

n00b question: Is it normal to get frame drops above about 85% GPU usage? I don't reach 90% at all but if I use any options that get me between 82-87% I get frame drops. :(

GTX 770, 347.88

I don't start dropping frames but I get a lot of presentation glitches when using settings that get above the high 80% GPU usage. I think this is normal. I seem to max out at ~95% usage and start dropping frames with settings that would use above that.

GTX Titan, 352.86

[ryrynz]

@madshi

I still prefer the high preset. Most of the time, there is no clear winner but I found an example where the high preset is clearly superior, you have to set very high values in Shiandow's debanding to be on par with madVR's debanding and by doing that you lose a lot of details.


Easy win there for f3kdb.

[Sunset1982]

I got some quesion for you guys:

I'm using MadVR, LAV and MPC-HC on an i5 4590 CPU + AMD R9 270 GPU.

Next week i'll get a 4K TV. So now I want to upscale my movies to 4K resolution.

Is this possible with good quality on my current hardware or do I need a better GPU?

What settings should I use to get best quality? Or should i try to get the best quality 1080p image out and let the tv do the upscale? (LG OLED)

[huhn]

I got some quesion for you guys:

I'm using MadVR, LAV and MPC-HC on an i5 4590 CPU + AMD R9 270 GPU.

Next week i'll get a 4K TV. So now I want to upscale my movies to 4K resolution.

Is this possible with good quality on my current hardware or do I need a better GPU?

What settings should I use to get best quality? Or should i try to get the best quality 1080p image out and let the tv do the upscale? (LG OLED)

24 hz 1080p can be upscaled with nnedi3 16 and nedi using a r9 270. i highly doubt a TV scaler can get close to this.

and don't forget forum rule 12: http://forum.doom9.org/forum-rules.htm

[ashlar42]

I seem to be unable to activate the D3D11 presentation mode.

I am on Windows 7 x64, madVR 0.88.8, LAV Filters 0.65.0 and XySubFilter 3.1.0.705.
dxdiag reports DirectX 11 as correctly functioning, GPU is GTX 660 with Nvidia drivers 352.86.

No matter the settings combinations I try on Rendering settings, I always get D3D9 with CTRL+J.
But if I choose the DX11 only dithering methods, they appear to be working just fine.

Anything worth checking?

[XMonarchY]

I tested low-mid 720p SMPTE 170M content I had with debanading algorithms and I'd have to agree with Werewolfy also. Shiandow's makes some places with low dark levels look smoother, but at the cost of removing detail. It also makes those smoothed-out areas stand out against the rest of the picture. So, for example, madVR's mid/high de-banding ends up showing noisy bands / compression squares in dark shadows, without taking out much detail and without making these noisy bands stand out. There is a rather smooth gradient transition from dark shadowy parts to brighter parts of the image. It feels natural - uniform. If you add or use only Shiandow's algorithm, that same dark shadow becomes really nice and smooth, but it loses detail and that part stands out against the rest of the image. It makes the dark shadow part look like 2-3 giant bands in my tests and on Werevolfy's posted image comparisons. It kind of makes it seems like the image uses / is of lower bit-depth...

I do think more tweaking could make it useful though. It needs to get better at 2 things - 1. preserving more detail at the cost of decreased smoothness and 2. merging smoothed out "banded" content with "non-banded" content better. Otherwise it simply creates fewer, smoother, but more obvious bands with less detail. If I had to pick what to fix - I'd rather fix #2 problem before #1 because loss of detail is not as much of a big deal for lower-quality content, but the look of smoothed out bands standing out against the rest of the image is quite annoying.

P.S. I feel SO bad critiquing other people's hard work...

[ryrynz]

24 hz 1080p can be upscaled with nnedi3 16 and nedi using a r9 270. i highly doubt a TV scaler can get close to this.


Actually LG's 4K sets do something very similar to NNEDI3.

[Shiandow]

I do think more tweaking could make it useful though. It needs to get better at 2 things - 1. preserving more detail at the cost of decreased smoothness and 2. merging smoothed out "banded" content with "non-banded" content better. Otherwise it simply creates fewer, smoother, but more obvious bands with less detail. If I had to pick what to fix - I'd rather fix #2 problem before #1 because loss of detail is not as much of a big deal for lower-quality content, but the look of smoothed out bands standing out against the rest of the image is quite annoying.

P.S. I feel SO bad critiquing other people's hard work...

I think the main cause for #2 is that MadVR's banding adds dithering, which tends to hide those "larger" bands a bit more. In theory what my debanding algorithm does is a bit closer to the source, but that's not much help if the source isn't good. Might be worth investigating if adding dithering helps, fortunately that's quite easy. It might also be possible to remove those larger bands entirely, but that would require reverse engineering how they were created which could be difficult, and might even be different for each file / frame, so that's not really something you can do with a general purpose algorithm.

Thanks for the criticism by the way, you were very clear in describing how and why it looked better / worse, which is very helpful, so there's no need to feel bad about it.

[Sunset1982]

Who else is using madvr with an amd card?

what settings should i use in Madvr and catalyst to get best results?

currently I'm using:

madvr: 16-235

AMD Controlpanel: RGB 4:4:4


is that ok or should I use something else?

[huhn]

Who else is using madvr with an amd card?

what settings should i use in Madvr and catalyst to get best results?

currently I'm using:

madvr: 16-235

AMD Controlpanel: RGB 4:4:4


is that ok or should I use something else?

i use a r9 270 my self.

all these settings depend on the screen and what the screen does with it.

in theory sending unlimited RGB to the GPU and outputting unlimited is the best way.

[iSunrise]

Actually LG's 4K sets do something very similar to NNEDI3.
I highly doubt it. How do you know? Did you do side-by-side comparisons?

Something along the lines of NNEDI3 is pretty wasteful of your GPU resources. And upscaling anything to 4K with NNEDI3 requires a lot of performance that usually only discrete desktop GPUs can handle.

This would be WAY too expensive for LG to put into their scaler unit, even for an ASIC that has largely fixed-function units, I really highly doubt they would go that far.

I have not seen real-world comparisons though, but if it's another variant of the so-called "Super Resolution" algorithms, I bet that not even Sony can deliver what madVR can do.

Feeding a 4K panel with very low-resolution content should be a good way to judge, but when you feed at least 720p or 1080p to it, differences will probably be very hard to see. Even the use of a very mild sharpener and basic upscaling suddenly makes Marketing-Departments scream "Super Resolution".

I have never owned a 4K display though and I haven't done comparisons with madVR on any of them, so consider this just an opinion, based on the way the monitor industry works, meaning they usually only implement small improvements, if at all.

[dbcooper]

i use a r9 270 my self.

all these settings depend on the screen and what the screen does with it.

in theory sending unlimited RGB to the GPU and outputting unlimited is the best way.

Do you have a TV that accepts RGB 0-255 at 24p? How about 4:4:4?

[huhn]

Do you have a TV that accepts RGB 0-255 at 24p? How about 4:4:4?

my TV accepts 4:4:4 at all refresh rates with unlimited range from 23p- 60p

[Sunset1982]

Do you have a TV that accepts RGB 0-255 at 24p? How about 4:4:4?

yes my tv accepts 0-255 at 24p.

It also accepts 4:4:4, but only in 60hz mode (desktop). when starting a movie and madvr switches to 23p/24p it doesnt get 4:4:4 anymore and changes automaticly.

picture seems good when using 0-255 in madvr and 4:4:4 RGB in Catalyst, when setting black to high in tv menu.

I just wonder what will be the best for quality to avoid any unnessecary convertion in madvr.

[MistahBonzai]

Do you have a TV that accepts RGB 0-255 at 24p? How about 4:4:4?

I'm using AMD HD-7850 with madVR/LAV/SVP and AviSynth+. Display is a run-of-the-mill 7 YO 40" Sony Bravia - it has 10-bit panel and supports full range RGB 4:4:4. CPU is intel i7-3770 with factory timing. I can still use the latest improvements to madVR by off-loading NNEDI3 and some filtering to the CPU (in AviSynth scripts via ffdshow raw video). You might be pleasantly surprised with what can be done with older hardware after investigating what it can do and spreading the load appropriately.

[Sunset1982]

@ madshi:


could you do an option to deactivate the border around the pattern in madtpg when using apl patterns?

This seems to be necessary to avoid autodimming on lg oled tv's. with the border around it the tv seems to dimm and destroys any calibration. without it I was told it should work.

[huhn]

yes my tv accepts 0-255 at 24p.

It also accepts 4:4:4, but only in 60hz mode (desktop). when starting a movie and madvr switches to 23p/24p it doesnt get 4:4:4 anymore and changes automaticly.

picture seems good when using 0-255 in madvr and 4:4:4 RGB in Catalyst, when setting black to high in tv menu.

I just wonder what will be the best for quality to avoid any unnessecary convertion in madvr.

try smooth motion at 60 HZ.

if you don't sit to close to the screen it should be good on a sample hold display. they don't have any motion resolution anyway.

[Sunset1982]

try smooth motion at 60 HZ.

thx for you suggestion. do I have any disadvantages using it instead of 24hz mode?

whats the good thing about it on sample and hold displays?

[huhn]

thx for you suggestion. do I have any disadvantages using it instead of 24hz mode?

whats the good thing about it on sample and hold displays?

the potential problems from smooth motion should be way less obvious on screens with high motion blur like lg OLEDS or sample hold screen in general without BFI /black frame interpolation.

in the end you have to find this out by your self.

if you don't see issue with active smooth motion at 60 Hz you will get 4:4:4 and smooth playback without frame creation interpolations artefacts.

when you use smooth motion make sure the screen interpolation is deactivated and that you didn't drop any frames what so ever.

smoothmotion needs a little bit of processing power and quiet a lot with error diffusion. on a well working screen with 60 Hz you should see any issue what so ever.

[Sunset1982]

thx. Tried it, but get repeated frames every 40 seconds. with 23hz mode i get repeated frames every 3 hours.

beside this, it looks good.

[huhn]

thx. Tried it, but get repeated frames every 40 seconds. with 23hz mode i get repeated frames every 3 hours.

beside this, it looks good.

the "repeat/drop every X" can be ignore with smooth motion. this is not going to happen

[dbcooper]

my TV accepts 4:4:4 at all refresh rates with unlimited range from 23p- 60p

Interesting. What TV is it? My Samsung only accepts both full range and 4:4:4 at 60 Hz.

[huhn]

Interesting. What TV is it? My Samsung only accepts both full range and 4:4:4 at 60 Hz.

an old cheap Phillips 4606 or something like that.

and it's totally normal that samsung can't do 4:4:4 at everything other than 60 hz and rarely on UHD screen at 30 HZ. a typical samsung flaw like nearly all other TV brands. they simply don't care.

[dbcooper]

Thanks, huhn. It really annoys me that my Samsung can't display full range inputs at non 60 Hz inputs.

[huhn]

Thanks, huhn. It really annoys me that my Samsung can't display full range inputs at non 60 Hz inputs.

i looking for years for a new TV but always these shortcuts left and right no 4:4:4 at all refresh rates not BFI with 4:4:4...

you could try smooth motion at 60 hz may work good on your screen may not find it out. GL

[XMonarchY]

I think the main cause for #2 is that MadVR's banding adds dithering, which tends to hide those "larger" bands a bit more. In theory what my debanding algorithm does is a bit closer to the source, but that's not much help if the source isn't good. Might be worth investigating if adding dithering helps, fortunately that's quite easy. It might also be possible to remove those larger bands entirely, but that would require reverse engineering how they were created which could be difficult, and might even be different for each file / frame, so that's not really something you can do with a general purpose algorithm.

Thanks for the criticism by the way, you were very clear in describing how and why it looked better / worse, which is very helpful, so there's no need to feel bad about it.

Your de-banding does work better on HQ HD sources better than on LW SD sources, but its the LW SD sources that NEED de-banding the most. Is there maybe a way of having madVR apply its own de-banding and dithering process on top /after YOUR de-banding algorithm, thus adding de-banding of de-banding + dithering on top of those larger smoothed out bands? I am not an expert on any of this, but AFAIK all processing is done one at a time, one on top of the other, etc. Another thought I had is - maybe to make those large bands gradually transition into other areas of the image you can blur / fade them? Again, I only understand the very surface of all these processes, so I don't know how viable these suggests are!

[Shiandow]

Your de-banding does work better on HQ HD sources better than on LW SD sources, but its the LW SD sources that NEED de-banding the most. Is there maybe a way of having madVR apply its own de-banding and dithering process on top /after YOUR de-banding algorithm, thus adding de-banding of de-banding + dithering on top of those larger smoothed out bands? I am not an expert on any of this, but AFAIK all processing is done one at a time, one on top of the other, etc. Another thought I had is - maybe to make those large bands gradually transition into other areas of the image you can blur / fade them? Again, I only understand the very surface of all these processes, so I don't know how viable these suggests are!

I don't entirely agree that only low quality sources need debanding, any source that is 8 bit (or maybe even 10 bit) needs debanding. Well, apart from those cases where they've added dithering before rounding to 8 bits, but those don't seem to be too common.

Anyway the problem with blurring those large bands is that it is impossible to distinguish them from actual detail. I suspect that those large bands are caused by the source skipping a value (e.g. instead of going 1,2,3,4, it goes 1,2,4,4). If that is the case you could fix / detect this by constructing histograms and rescaling the values to mask the "missing" ones, but that would require analysing the entire source, so I don't think it's something you'd want to do in real time.

[shaolin95]

Guys I am trying to understand the OSD. Do you see anything wrong here?
This was just testing the 10bit display option so no movie playing but if someone can tell me if you see anything wrong and how to interpret all those things listed there, I will really appreciate it! :)
BTW, I had ffdshow (I am using it for SVP) set to YUV12 instead of Auto of RGB. Not sure if that was how SVP set it up automatically but brings up the question on which option I should be using for the output in ffdshow filter.
For Audio and video I am using LAV filters.

Another Q about madvr...ever since I checked the WASAPI option, I am unable to remove it. I uncheck it but never goes away.

In any case, I am using a GTX 750ti and have the Nvidia settings as RGB and 0-255.
Any recommendations on maxing out IQ will be appreciated. :)

http://i59.tinypic.com/4vobqv.jpg

[huhn]

Guys I am trying to understand the OSD. Do you see anything wrong here?
This was just testing the 10bit display option so no movie playing but if someone can tell me if you see anything wrong and how to interpret all those things listed there, I will really appreciate it! :)

it can take some time before we can see the picture. think about an external image host.

BTW, I had ffdshow (I am using it for SVP) set to YUV12 instead of Auto of RGB. Not sure if that was how SVP set it up automatically but brings up the question on which option I should be using for the output in ffdshow filter.
For Audio and video I am using LAV filters.
leave all colorspaces active so it can choice the one that needs no conversation.

Another Q about madvr...ever since I checked the WASAPI option, I am unable to remove it. I uncheck it but never goes away.
what has madVR to do with wasapi?

[shaolin95]

it can take some time before we can see the picture. think about an external image host.


leave all colorspaces active so it can choice the one that needs no conversation.


what has madVR to do with wasapi?

Ok so I selected all options in the ffdshow filter option and then set it to Auto.
I added a link to the picture now.
Also, sorry I mean mpc not madVR about the WASAPI issue but that question really does not belong here so feel free to ignore it.
Thanks

[huhn]

Ok so I selected all options in the ffdshow filter option and then set it to Auto.
I added a link to the picture now.
Also, sorry I mean mpc not madVR about the WASAPI issue but that question really does not belong here so feel free to ignore it.
Thanks

it's a standing picture not a lot to judge from. i don't know why you use lanczos 8 for everything but do as you pleases.

the 16 bit test picture doesn't work with ffdshow and is only inputted as 8 bit so if you judge 10 bit this way you just lied to your self X-)

[shaolin95]

it's a standing picture not a lot to judge from. i don't know why you use lanczos 8 for everything but do as you pleases.

the 16 bit test picture doesn't work with ffdshow and is only inputted as 8 bit so if you judge 10 bit this way you just lied to your self X-)

This is Avengers without ffdhow (I normally dont use this configuration but just to get an idea of what my performance is and what the numbers mean):
http://i58.tinypic.com/1ili0n.jpg

And this is Avengers with ffdshow and SVP (hence I changed from NEDI to Lanc to save processing power):

http://i60.tinypic.com/2podxqc.jpg


I did the test without ffdshow raw video filter and with it and both times I can clearly see the banding at 8 bit but not when on 10bit.
No idea why but that is what my system shows.
In any case, I may go back to just not using SVP but its such a toss up...either way I am dealing with artifacts in motion or added by SVP.
Thanks for your help!

[huhn]

not sure why you switch from a neutral resizer like nnedi3 to a totally over sharpening and bad resizer like lanczos 8 AR and not something like lanczos 3 ar or jinc 3 ar. more tabs on such resizer is not better.

[shaolin95]

not sure why you switch from a neutral resizer like nnedi3 to a totally over sharpening and bad resizer like lanczos 8 AR and not something like lanczos 3 ar or jinc 3 ar. more tabs on such resizer is not better.

because this is precisely what I am trying to do with this post, learn.

And also learn what all those extra things mean like max stats, average stats etc so when I am making some adjustments I can tell how bad it is affecting my performance (aside from obvious lag of course).
I am sure there must be some article or something explaining this but so far not finding exactly what I am trying to learn.
Thanks

[huhn]

this is the best place to start.
http://forum.doom9.org/showthread.php?t=171787

[ryrynz]

I highly doubt it. How do you know? Did you do side-by-side comparisons?


I know because I have a number of 4k sets here at work at there's no way those edges I'm seeing on content is not the result of something similar to NNEDI3 if not in source at least in appearance. I can show you screenshots if you wish so your doubt is alleviated. When you've got that many extra pixels to work with it just makes sense really, wasted opportunity otherwise. Most notice able on the edges of letters where full hd screens show rather jagged edges and the ultra hd ones are perfectly smooth on < high def content.

[mithra66]

I'm actually struggling to make MadVR 10 bits mode work with custom 23Hz resolution on NVidia GTX970.

But is it worth it ?

Pioneer LX6090 offically support deep color (36 bit), but I don't know how it manages it. Is it true 10 bits or some dithering with lower quality than MadVR 8 bits dithering?

What would be the best setting for the 9g Kuro display? 10 bit output, 8bit or even 6 or 7 bit output with MadVR dithering , depending on the panel output capability?

[6233638]

I don't entirely agree that only low quality sources need debanding, any source that is 8 bit (or maybe even 10 bit) needs debanding. Well, apart from those cases where they've added dithering before rounding to 8 bits, but those don't seem to be too common.

Anyway the problem with blurring those large bands is that it is impossible to distinguish them from actual detail. I suspect that those large bands are caused by the source skipping a value (e.g. instead of going 1,2,3,4, it goes 1,2,4,4). If that is the case you could fix / detect this by constructing histograms and rescaling the values to mask the "missing" ones, but that would require analysing the entire source, so I don't think it's something you'd want to do in real time.I wish I had the time to dedicate to testing your debanding algorithm now, but I agree that any 8-bit source probably does benefit from debanding.
Even if the source was dithered to 8-bits or has film grain, video compression is going to add some degree of banding back into the source.

I actually had a very difficult time finding a true 8-bit gradient with 256 steps that were evenly spaced, so I had to create my own.
I think most people seem to have just used a gradient tool to create a gradient from black to white that was 256px wide without actually checking it.

Here is a 256px Gradient (http://abload.de/img/169-gradientkarnq.png)
And here it is pre-scaled for 1080p without any dithering (http://abload.de/img/1080p-gradient0jqpk.png)

Assuming the image host has not modified the images (some hosts re-compress them) these should be "perfect" undithered 8-bit gradients.
What should be noticeable is that there is evenly spaced banding (every 7px) in the large image.
If there is any unevenness, that is likely due to your display's calibration, or it not being able to display a true 8-bits.

But the point is that there is clear banding with a true 8-bit gradient if there is no dither/noise to mask it.
Using some form of debanding will greatly improve the smoothness of that image even though it's a "perfect" 8-bit gradient with evenly spaced steps.

I'm actually struggling to make MadVR 10 bits mode work with custom 23Hz resolution on NVidia GTX970.
But is it worth it ?
Pioneer LX6090 offically support deep color (36 bit), but I don't know how it manages it. Is it true 10 bits or some dithering with lower quality than MadVR 8 bits dithering?
What would be the best setting for the 9g Kuro display? 10 bit output, 8bit or even 6 or 7 bit output with MadVR dithering , depending on the panel output capability?Plasmas are natively 1-bit displays which use a whole lot of dithering to create shades.
Now that doesn't mean you should send them a 1-bit signal from madVR, but I would be surprised if it made a bit of difference whether you send them an 8-bit or a 10-bit input.
Perhaps the last generation or two of Panasonic sets might benefit from it, but I have my doubts.
Pioneer sets almost certainly will not. They have the fewest steps of gradation of any plasma model due to them prioritizing the elimination of dynamic false contouring and black level over all else.

I also wonder if it might be worthwhile to send them 60Hz and use SmoothMotion, because 60Hz is a marked improvement for gradation over 24/50Hz inputs on those displays.
The higher the refresh rate, the fewer steps of gradation a plasma has, and Pioneer displays those inputs at 72/100Hz.

[TobiMan]

Here is a 256px Gradient (http://abload.de/img/169-gradientkarnq.png)
And here it is pre-scaled for 1080p without any dithering (http://abload.de/img/1080p-gradient0jqpk.png)


:thanks: Pretty handy for testing!

[mithra66]

Thank you 6233638 for your reasoned answer :)
For the moment I think I prefer 12 bits output to 8 bits dithered. But it may be placebo effect ;-)
I will proceed more testing .

[madshi]

Maybe both algos can be combined at "ultra"?
There is hardly an additional detail loss when I enable old high and Shiandow's at 0.1/0 at the same time, but Shiandow's algo clearly reduces banding left by madVR's old one.
Might be an option, but running both algos also means double GPU power. So it'd be better to try to tweak one of the algos so that most people consider it better than the other.

i have no modes listed there, i use the automatic detection. i think...
Oh, that's a weird combination, telling madVR to switch display modes, but not giving it any modes to switch to. That's not the way it's supposed to be used. You should enter the list of modes you want madVR to switch to. If you leave the mode list empty, you can just as well disable the whole option.

anyway, when i go into fullscreen exclusive mode usually what happens is that it displays 60hz/60fps also in OSD.

And thats what im used to, works perfect..

now when i enable this d3d11 mode, it goes to 25hz, changes the system resolution to TRUE 1920x1080 @ 25hz, wich is not what should happen, i guess.. and therefore teh picture isnt as fluid anymore..
Does any of this change if you provide madVR with a list of modes you want to switch to? Or if you totally disable the madVR refresh rate switching?

Is there a methodical way to determine whether flush settings should be used and which ones? I have disabled flush on my GTX 670 yet it seems my GT 750M works better with the default flush, flush/sleep, no flush, no flush settings.
I'm always using the defaults. I'd recommend to always start there. If you've no problems, just keep the defaults. Maybe tweaking the settings could result in a small performance improvement sometimes, but it also risks new problems.

Same here
10bits FSE gives me 4-5 rendered frames, with any settings. Instead with D3D9 I can go up and fill my queue.
You could try lowering the number of prepresented frames step by step. It's possible that your GPU driver can't handle as many prepresented frames when using 10bit output.

Does this also hold true while having smooth motion activated?
Depends on the exact movie fps and display refresh rate, but for e.g. 24fps @ 60Hz it also holds true.

2) In exclusive mode settings, I have how many frames shall be presented in advance at 3. What means this options?

3) What means cpu queue size and gpu queue size? Is it necessary to modifiy the defaults even if no problem (12 forcpu and 8 for gpu)?
3 in advance is a bit low. Usually the defaults should be fine. Increasing the queue sizes will increase CPU/GPU RAM consumption. But if you have lots of RAM to spare, increasing the GPU sizes could improve protection against frame drops if you PC gets "busy" for short periods of time.

madshi, am I correct in assuming you have no intention of keeping both madVR's and Shiandow's debanding algorithms?
Correct!

I understand that FSE mode is required to get a 10-bit output. However, is it necessary to use a DX11 windowed mode?
I still prefer the old windowed mode since it responds immediately to inputs, instead of the delay that you get from presenting several frames in advance.
I might find a way to improve that for DX11 windowed mode in the future. Or maybe not. But yes, the DX11 FSE mode requires DX11 windowed mode to be used - at least in the current madVR build. In theory I could also mix that. But it would make my code more complicated, and there are already so many different "paths". So I think I prefer to keep it the way it is. Furthermore, there are rumors that Windows 10 might add support for 10bit output in windowed mode. Not sure if that's true, but in any case, I don't plan to offer the combination of D3D9 windowed + D3D11 FSE atm.

When I compare 8-bit and 10-bit, there is a clear improvement.
An improvement in smoothness or in noise? Of course outputting 10bit means that madVR adds a lower amount of noise to the image. That might be what you see as an improvement. But if your display doesn't really handle 10bit well, this would come at the cost of a loss in gradiant smoothness.

However, with 8-bit using Error Diffusion I was actually seeing that some tests seemed to be smoother in 8-bit rather than 10-bit.
This could be attributed to the dithering algo (error diffusion vs ordered dithering), but it could also mean that your display doesn't handle 10bit input as well as it should. Might require some tests to find out which is which, once I've allowed error diffusion to be used for 10bit output again.

Is there any reason to still include madLevels Tweaker? nVidia now has a fully functional setting that provides full range 0-255 over HDMI in videos and games. Is there any reason to use it?
It's still there because not everybody might want to use the latest drivers. Also it might still be useful for Intel users. I don't know.

Another question I had - does MPC-HC override madVR refresh rate settings? For example, my TV supports 23hz and 24hz modes, but they are not exactly 23.976hz because of my videocard. If I create a custom 23.976hz mode in MPC-HC x64 1.7.8 build 191, then will that 23.976hz apply or will madVR just use my TV's default 23hz mode? I also tried to make a 23.976hz custom resolution in nVidia CP, but it creates a 24.000hz resolution instead. How can I force it to use 23.976hz exactly?
You should either use MPC-HC for refresh rate switching or madVR, but not both. If you use madVR's refresh rate switching, and disable the MPC-HC one, the MPC-HC options won't have any effect on madVR.

I am using latest madvr version.I am on windows 10,nvidia gtx 980 latest version(windows 10 driver)
I am not able to switch to NNEDI in chroma upscaling or use image doubling.It crashes.Can anybody confirm or look into it.
I am using nvidia version 352.84. I am using pot player 64 bit
Latest NVidia drivers are hit and miss with NNEDI and error diffusion. You may have to try different driver versions to find one which works.

Though the 10-bit output was working, I had to reduce this from 16 to 12 on my system to stop the queues being almost empty in the 10-bit DX11 FSE mode.
Only 10-bit was affected, 8-bit DX11 FSE worked just fine at 15. (16)
I noticed this same issue last night, with v0.88.8. I usually use lower buffers of 6 present and 8 render but I was experimenting with turning them up again. As soon as I go over 8 present the rendering queue gets stuck at 1-4 full and present stays similarly low. No dropped frames or anything though. With 8 or below presented in advance I can use 24 rendering buffers without any issues.
Interesting! Seems that your GPU driver doesn't support as many prepresented frames when 10bit output is active.

Can you confirm that the madVR default settings for the queue sizes and prepresented frames (but with DX11 activated, of course) work fine?

I may be wrong, maybe madshi will be able to call 10bit in windowed mode if the desktop is already in 10bit...
No. As long as Windows itself (DWM/Aero etc) is 8bit, there's no use for me to output 10bit in windowed mode. It would just introduce banding. Maybe things will change with Windows 10, but I don't know.

FWIW I'm still getting 23.969 refresh rate whenever D3D 11 is on (ie FSE and D3D both enabled), even with 88.8 and the 352.86 nvidia driver. Doesn't matter if I set madvr & nvidia to either 8 or 10 bit, same result. I gather from the discussion earlier this is an issue with D3D 11, but just wondered if anyone else was seeing this. D3D 9 gives me 23.976 just fine.
I get 23.970 on my GTX 970 with D3D11.

I have a custom resolution set-up for nearly perfect 23.976 in D3D9 but it seems to be getting overridden in D3D11 exclusive mode for some reason.
There'll be some changes in the next build, maybe that will improve things, but I'm not sure.

For double videocards users, how to force one of them to render?

On Win8.1 64, HC64 + madVR64, both DX9 and DX11, why can't get into fullscreen exclusive mode?
It might be possible to choose one for OpenCL work, but not for normal DX9/11 rendering. madVR asks the OS which GPU is responsible for the current screen and uses that.

Not being able to go to FSE mode could suggest that the "wrong" GPU is used, but I'm not sure. I've 3 different GPUs in my development PC (one each Intel, NVidia and AMD) and I can go FSE with every GPU just fine. The OS correctly reports the correct GPU for each HDMI/DVI/VGA output.

this video renderer lives its own life ;)
before i had fullscren exclusive mode at 10 bit
now the it says "fullscreen exclusive" NEW PATH

what is that?
and why the does it change settings on its own?
i havnt changed anything.
Which madVR version are you using? Make sure you're using v0.88.8. With some previous build 10bit output failed if you had the number of prepresented frames set to 16.

There's a new superresolution upscaler for still images: https://github.com/nagadomi/waifu2x/ based on http://arxiv.org/abs/1501.00092

There also is a webservice running that implementation which has been trained with anime-style images and provides very good results on that type of material, especially on fine lines that are barely a pixel wide and fairly aliased. It also seems to introduce fewer fake details on flat surfaces than nnedi.
If you run it on real life images you'll get oil paintings.
Specialization a specific kind of content seems to have a pretty big impact.

Source image (http://i.imgur.com/uWMu9zt.png)
new upscaler w/ noise reduction (http://i.imgur.com/WnHLPE4.png)
new upscaler w/o noise reduction (http://i.imgur.com/QG8aVtr.png)
NNEDI3 32 w/ SuperRes (http://a.pomf.se/ywwfia.png)
NNEDI3 64 w/o SuperRes (http://a.pomf.se/smpgig.png)


I figure it's too expensive to run this in realtime, thus my question is can we get madvr any closer to those results? E.g. by having separate live action / animation modes for nnedi3?

Edit: added needi w/o superres
It does indeed look very good. But as Shiandow already said, this is probably dramatically slower than NNEDI3. Furthermore, I've just tested it with some of my usual upscaling test images and in one image this new upscaler produced a very weird and ugly image artifact. I guess in order to use this, a much bigger image database would have to be used for training first, to make it really reliable and get of artifacts. Still, it would be terribly slow, unfortunately...

@Madshi: About the fix you wrote for Windows 8 59/60 23/24 Hz auto refresh rate change. It works well for my ATI HD 5770, but with Intel HD 2000, it can't prevent a quick change in refresh rate. I mean

Without your fix, with Windows 8 bug
59 Hz -> no changes
60 Hz -> 59 Hz

With your fix on ATI 5770
59 Hz -> no changes
60 Hz -> no changes

With your fix on Intel HD 2000
59 Hz -> no changes
60 Hz -> quick change to 59 Hz -> quick changes to 60 Hz, each time I enter FullScreen (exclusive)
There's probably nothing I can do about it, unfortunately.

Does DX11 provide a performance improvement over overlay mode?
I don't think so.

I downloaded a 60fps video from YouTube and it skips back and forth with DirectX 11. SVP is off.
The skipping back & forth is an Intel driver problem. Nothing I can do about it, unfortunately... :(

a small bug report(if it can even be called a bug).the batch files for install & uninstall are not working on my pc.i have to manually use command prompt to register madvr filter.it worked last time but not now.running install batch file as admin simply result in a 1sec appearance & disappearance of cmd window.
<sigh> This is proving to be much more difficult than I thought. I'll try to modify the batch file once more for the next build. If that fails, I might have to make it dumber / less comfortable.

I don't know what the heck happened but I'm getting ridiculous rendering times using NNEDI3 after updating to the latest version of madVR. 20ms higher render times on my AMD 7850.

I'm going to restart my computer to see if something got messed up.

Edit: Nope. I'm at a loss for words. I haven't changed anything on my end except to continue to update Windows 8.1.
Did you try with and without the "use alternative interop hack" option?

I seem to be unable to activate the D3D11 presentation mode.

I am on Windows 7 x64
Do you have the Windows 7 Platform Update installed?

I actually had a very difficult time finding a true 8-bit gradient with 256 steps that were evenly spaced
Try the first post of this thread. "madTestPatternSource" is what you want.

could you do an option to deactivate the border around the pattern in madtpg when using apl patterns?

This seems to be necessary to avoid autodimming on lg oled tv's. with the border around it the tv seems to dimm and destroys any calibration. without it I was told it should work.
According to the latest posts on AVSForum that's not true. autodimming on LG OLED TVs seems to be avoided just fine with madTPG as it currently is. Or do you have a different experience? I think the AVSForum is the better place to discuss that, though:

http://avsforum.com/forum/showthread2.php?t=1471169

When D3D11 path is activated, D3D11 debug runtime reports some objects leaking. I think if you create device with D3D11_CREATE_DEVICE_DEBUG flag you should see them in debug output
I've tried, and I see some DXGI messages, but no complaints from D3D11. Do I need to do something specific to make these appear?

[madshi]

As far as debanding goes I actually prefer Shiandow's over MadVR's for anime; and madVR's over Shiandow's for movies and DVR files. Shiandow's is more aggressive and kills some details, which isn't really that big of an issue for anime imho. MadVR's preserves details better while still doing a very good job of overall debanding which for me, works better for movies. This is just my opinion and isn't based on any thorough concrete testing so "take it with a grain of salt".
After some training I learned to clearly differentiate between the two debanding algos.

Some findings about Shiandow's algo:
Shiandow's default settings (0.50, 0.00) are somewhere between Mid and High of madVR in terms of debanding efficiency.
Retains general shape better than madVR, hence less "detail" loss.
Smoother with less dithering/grain visible (I see tiny squares [2x2 pixels] with madVR debanding).

The blending between gradients of madVR debanding adds clearly visible grain to the transitions and retains less of the general shape.
On the opposite, in Shiandow's I can't see the grain (silky smooth) and it retains the general shape a lot better.
In this iteration of madVR (0.88.8) Shiandow's is a clear winner compared to Mid and High.

In Low setting it's the opposite (quick [not thorough] test).
I like madVR's Low better than any other setting of Shiandw's (margin on 0.00).
This grain that removed the general shape on Mid and High is important part of Low.
I did not find any setting on Shiandow's that does a better job than Low.
This without playing with Margin which I don't see if it works at all.
As far as “low” and “medium” presets are concerned, the original debanding algorithms seem to be perfect. I clearly prefer them over Shiandow’s debanding.

For “high” preset I am not sure which one is better. However, I found one case of a poorly processed video where the old “high” preset worked perfectly but Shiandow’s algo could not be tuned to remove banding that much effectively. Please see the screenshots in the links below. I would maintain that the old “high” preset should be tuned if required but not be replaced.

Original Image
http://www.mediafire.com/view/nst4vrzmaope170/original.png

Shiandow 0.1, 0.0
http://www.mediafire.com/view/jyauw3xnxz5q8n8/Shiandow%200.1%2C%200.00.png

Shiandow 0.4, 0.01
http://www.mediafire.com/view/xe6apy95mf5bz6a/Shiandow%200.4%2C%200.01.png

Shiandow 0.6, 0.03
http://www.mediafire.com/view/844c85asacq3kew/Shiandow%200.6%2C%200.03.png

madVR old “high” preset
http://www.mediafire.com/view/pdfipkv918aacts/madVR%20old%20high%20preset.png

Having stated my preference, I should also mention that after the new changes being made, Shiandow’s debanding has become really very good. There is nothing to complain, it’s just that in comparison I am preferring the old presets.
Regarding Debanding - I can't say one is better than the other. Just when I think one is better, I switch to a different source to see that my preference has changed. My personal preference is "less is more" so I generally prefer the lowest setting. I'd rather see a bit of banding than a washed out scene. My "vote" goes to the more efficient algo which appears to be MadVR (according to GPU-Z).
After some tests in the previous days, I came to the conclusion that I have no conclusive opinion. MadVR's algo does a fine job at most cases, then there are some where Shiandow's is better. The same could be said for detail preservation, thou MadVR's high felt a bit worse overall in this regard. Overall, Shiandow's algo matches MadVR's in its latest iteration (At the beginning I preferred MadVR's, but it improved a lot. Good job, Shiandow!). If anything, the "High" preset should be tested against Shiandow's more, but I'm afraid this is all I could do.
I tested the latest version of Shiandow's debanding and I still prefer the high preset. Most of the time, there is no clear winner but I found an example where the high preset is clearly superior, you have to set very high values in Shiandow's debanding to be on par with madVR's debanding and by doing that you lose a lot of details.

High preset : [img=http://s5.postimg.org/6enu7he9z/high1.png] (http://postimg.org/image/q99vtlthf/full/)
[/url]
Shiandow - power : 0.80 - margin : 0.00 : [img=http://s5.postimg.org/qnb7t7dl3/shiandow085000.png] (http://postimage.org/index.php?lang=french)

Shiandow - power : 0.85 - margin 0.30 : [img=http://s5.postimg.org/9ckt16lxj/shiandow085030.png] (http://postimage.org/index.php?lang=french)

Shiandow - power : 0.90 - margin : 0.50 : [img=http://s5.postimg.org/ushj1dlzb/shiandow090040.png] (http://postimage.org/index.php?lang=french)



Sometimes Shiandow's algo is better, sometimes it's weaker. MadVR's debanding smoothes the banding and Shiandow's algo blurs it. It also blurs compression artefacts which could be usefull but it's not really the role of a debanding algo and I think that a denoise or a deblocker will be a lot more efficient at this task.

I also played with the high preset and I found a way to improve it a little bit by setting Angleboost at 1.4 and maxAngle at 0.10.
You don't lose more details than with the original high preset and it can help sometimes.

[img=http://s5.postimg.org/gnbpzkcxz/highc1.png] (http://postimage.org/index.php?lang=french)
[url=http://postimage.org/index.php?lang=french] (don't pay attention to avgDif, it's a mistake and it dosen't change anything)

I know that the difference is not huge but you don't lose more details by doing that so I don't really see why not doing it ;) With these settings, I find hard to find an example where Shiandow's algo is better.
I must agree with Werewolfy. After switching between the algo's for a few days I've come to the conclusion that madVR's high preset does a better job with less blurring.
I have wanted to provide feedback on debanding but my experiments have failed to provide useful distinctions (unless a report of inconclusive is useful). :o

Depending on the source and the exact banding and details I am paying attention to I can sometimes notice differences but I have not found anything that would allow me to feel comfortable saying generally one is better than the other.

I generally use no or light debanding, conceptually I prefer viewing banding to losing details. I haven't done any blind tests to determine what I actually prefer. :devil:
I tested low-mid 720p SMPTE 170M content I had with debanading algorithms and I'd have to agree with Werewolfy also. Shiandow's makes some places with low dark levels look smoother, but at the cost of removing detail. It also makes those smoothed-out areas stand out against the rest of the picture. So, for example, madVR's mid/high de-banding ends up showing noisy bands / compression squares in dark shadows, without taking out much detail and without making these noisy bands stand out. There is a rather smooth gradient transition from dark shadowy parts to brighter parts of the image. It feels natural - uniform. If you add or use only Shiandow's algorithm, that same dark shadow becomes really nice and smooth, but it loses detail and that part stands out against the rest of the image. It makes the dark shadow part look like 2-3 giant bands in my tests and on Werevolfy's posted image comparisons. It kind of makes it seems like the image uses / is of lower bit-depth...

I do think more tweaking could make it useful though. It needs to get better at 2 things - 1. preserving more detail at the cost of decreased smoothness and 2. merging smoothed out "banded" content with "non-banded" content better. Otherwise it simply creates fewer, smoother, but more obvious bands with less detail. If I had to pick what to fix - I'd rather fix #2 problem before #1 because loss of detail is not as much of a big deal for lower-quality content, but the look of smoothed out bands standing out against the rest of the image is quite annoying.
Thanks for your feedback, everyone!

The next build will contain another update from Shiandow again. Also I'll slightly modify madVR's "high" preset, based on Werewolfy's suggestions. So might I ask you guys to test again with the next build? Hopefully it will be the last debanding test build. After that I would like to make a final decision. Thanks!