madVR Options Explained

[Asmodian]

Another new version! I need to stop all this traveling!

Awesome news. :)

[cyber201]

@Asmodian
Yes, i need 10bit because 4k HDR files are in 10bit.

[huhn]

that's not how it works.there are nearly no true 10 bit displays out there. and 8 bit is fine.

if you take heavy compression and dithering into account than 10 bit isn't worth much for presentation but a lot for the compression.

[XMonarchY]

Is it possible to take screenshots in madVR in Exclusive FullScreen Mode?

[huhn]

there are ways to do that but this can easily cause playback issues.

[XMonarchY]

there are ways to do that but this can easily cause playback issues.

As long as it does not mess with Stats Info and the issue goes away after you restart the player or even PC - I am all for it, otherwise I will have to take photos...

[Asmodian]

I don't know how and even if you can causing playback issues would mess with the stats info so photos might be the best option... :p

[Mainge]

Hello Asmodian
I have a question, which I hope is not too stupid.
I just built a new PC with GTX1070 AMP Extrem. With all that new power, I can finally use MadVR has it should and I put the highest possible settings that my GPU can take (NGU sharp very high for Luma and so on). In addition, I also put direct supersampling on always doubling and I even tried direct quadrupling.

For information, I almost only watch 1080p source on 1080p display, which is what leads to my question.

In some of your posts, I could see that you don't recommand to supersample when source matches the display.
Should I really remove supersampling option? What is the downside of Upscaling to 2160p and then downscaling to get back to 1080p? I honestly thought this was the way to go in order to get super picture quality.

I am the type to put the options on the highest level when I can, but I know it can be totally wrong to do that...

Thanks in advance for your feedback.

[Asmodian]

Updated for madVR v0.91.10.

Added a description for a new option for the deep hack in Direct3D to enable true 24 Hz and 60 Hz modes in Windows 8.1 and 10. This option used to be always enabled but it sometimes causes the presentation queues to not fill.

[nsnhd]

By playing some hdr files and switch on madVR OSD, some files show "matrix BT.709, primaries BT.2020" and others show both "matrix & primaries BT.2020".
What that means and what's the difference ?

[Asmodian]

Matrix is the set of weights used to covert YCbCr data to RGB and primaries are the color of 100% Red, Green, and Blue. You can think of the matrix as the way to covert to RGB and primaries as the gamut of the RGB space once the YCbCr has been converted.

[nsnhd]

Matrix is the set of weights used to covert YCbCr data to RGB and primaries are the color of 100% Red, Green, and Blue. You can think of the matrix as the way to covert to RGB and primaries as the gamut of the RGB space once the YCbCr has been converted.

So, if matrix is BT.709 then is the file a real HDR one, or just a SDR with garmut BT.2020 added ?

[Asmodian]

There are a lot of real HDR videos with BT.709 matrix, it is simply the values used to multiply Y, Cb, and Cr by (while summing the results) in order to calculate R, G, and B. Mastering happens in RGB (YCbCr data is not even viewable as a color image) and as long as you use the correct inverse matrix when converting from YCbCr as you did when converting to it you get the same RGB values back.

The primaries are the important thing, anything with BT.2020 or DCI-P3 primaries could be real HDR, but you cannot tell from just that. How it was mastered and the other metadata is also important, but the matrix used to convert to YCbCr isn't. There will be differences in the rounding errors when converting between RGB and YCbCr with a different matrix; BT.2020 is probably theoretically better for 10+ bit HDR content for some reason(?), but not noticeably so.

[Asmodian]

Updated for madVR v0.91.11. I only added some extra description of the current state of HDR support in Windows 10 and a link to madshi's detailed description (https://forum.doom9.org/showthread.php?p=1810891#post1810891). I updated my settings.bin as well.

This version was mostly fixes for HDR support. :)

[Asmodian]

Hello Asmodian
I have a question, which I hope is not too stupid.
I just built a new PC with GTX1070 AMP Extrem. With all that new power, I can finally use MadVR has it should and I put the highest possible settings that my GPU can take (NGU sharp very high for Luma and so on). In addition, I also put direct supersampling on always doubling and I even tried direct quadrupling.

For information, I almost only watch 1080p source on 1080p display, which is what leads to my question.

In some of your posts, I could see that you don't recommand to supersample when source matches the display.
Should I really remove supersampling option? What is the downside of Upscaling to 2160p and then downscaling to get back to 1080p? I honestly thought this was the way to go in order to get super picture quality.

I am the type to put the options on the highest level when I can, but I know it can be totally wrong to do that...

Thanks in advance for your feedback.

I am not sure why I missed your post, I am sorry for that. :(

I don't like pure supersampling, it tends to introduce the minor artifacts that are the result of any scaling technique without any benefit. At best even NGU upscaling looks basically like the source after downscaling to the source resolution again. The source was probably downscaled from a higher resolution when it was mastered after all and even NGU upscaling isn't going to improve it beyond that original high resolution master.

The only case where I might use supersampling was if I was using strong upscaling refinements and I liked the look better than simply using image enhancements.

[Mainge]

I am not sure why I missed your post, I am sorry for that. :(

I don't like pure supersampling, it tends to introduce the minor artifacts that are the result of any scaling technique without any benefit. At best even NGU upscaling looks basically like the source after downscaling to the source resolution again. The source was probably downscaled from a higher resolution when it was mastered after all and even NGU upscaling isn't going to improve it beyond that original high resolution master.

The only case where I might use supersampling was if I was using strong upscaling refinements and I liked the look better than simply using image enhancements.

Thank you for your answer! Everything you said makes sense, I didn't know that the source is usually downscaled from higher resolution, you learn everyday :-)
So I think I'll remove this supersampling and try to play with Image enhancement instead of Upscaling refinements.

Thanks again!

[Manni]

Hi Asmodian,

Great thread, thanks for this work.

I have a couple of questions:

1) I just got a 1080 TI (EVGA cs2) like yourself so I'm going to take a look at your settings.bin. I'm coming from an AMD HD7870 and all my settings were optimized for a much less powerful AMD GPU. Do you have a profile for 3D in your settings?

2) One thing I'm not sure about, I set LAV HW acceleration to none because my understanding is that in that case MadVR is handling everything, not using the GPU h/w acceleration. For a max quality result, is it what I should use, or is there no drawback using DXVA2 copyback?

3) I'm abroad at the moment and had issues when I first tested with 3D (using mk3D files made with makemkv from my 3D Blurays). AMD handles 3D automatically while nVidia needs to enable 3D in the OS. I don't want 3D on al the time, especially when playing 2D video, so what's the right combination to use in the nVidia control panel settings and in MadVR to get 3D enabled automatically when playing 3D, and disabled afterwards? I tried the option to restore the OS setting but the switch back and forth in and out of 3D wasn't working 100% of the time.

In case it matters I set the resolution to UHD to upscale 2D blurays, but not 3D blurays. My display is a JVC rs500 projector. I play most of my UHD Bluray / HDR content with my Pana UB900, so will play with HDR when the nVidia HDR bug in FSE will be fixed (as we need FSE for 10bits, the JVC supports 12bits from the input to the panels).

Until we can get legt UHD Bluray content onto HTPC, my main use of MadVR is 2D/3D bluray playback (film mostly).

Thanks!

[Asmodian]

1) I do not have 3D profiles configured, 3D is actually disabled in my settings.bin. I do have a 3D display but I have not used 3D in a few years and I have not tested it at all. Sorry. I might test it again if people are interested but I thought it wasn't harder to display than 1080p60?

2) DXVA2 copyback offers the same quality video as software decoding but software can sometimes handle oddities better and/or is more responsive. Personally I use DXVA2 copyback only for UHD. Playing UHD HEVC or AVC takes too much cpu power and Pascal GPU's hardware decoding is very capable.

3) Note my answer to 1; sorry but I don't have any recent experiance with 3D playback with madVR. Hopefully someone else will offer some advice. :)

You actually do not need FSE for 10-bit anymore if you enable Windows' HDR mode. That said 10-bit is not important due to madVR's high quality dithering. Have you done any tests of 10-bit v.s. 8-bit output? I have never seen any differences with dithering enabled (it should never be disabled) on any 10-bit displays I have tested it on. 10-bit or higher is very beneficial for compression and processing but not for the final display, assuming good dithering is done. If you don't have to sacrifice anything for it 10-bit is good but once you start making any sacrifices you would probably be better off using 8-bit instead.

[Manni]

1) I do not have 3D profiles configured, 3D is actually disabled in my settings.bin. I do have a 3D display but I have not used 3D in a few years and I have not tested it at all. Sorry. I might test it again if people are interested but I thought it wasn't harder to display than 1080p60?

2) DXVA2 copyback offers the same quality video as software decoding but software can sometimes handle oddities better and/or is more responsive. Personally I use DXVA2 copyback only for UHD. Playing UHD HEVC or AVC takes too much cpu power and Pascal GPU's hardware decoding is very capable.

3) Note my answer to 1; sorry but I don't have any recent experiance with 3D playback with madVR. Hopefully someone else will offer some advice. :)

You actually do not need FSE for 10-bit anymore if you enable Windows' HDR mode. That said 10-bit is not important due to madVR's high quality dithering. Have you done any tests of 10-bit v.s. 8-bit output? I have never seen any differences with dithering enabled (it should never be disabled) on any 10-bit displays I have tested it on. 10-bit or higher is very beneficial for compression and processing but not for the final display, assuming good dithering is done. If you don't have to sacrifice anything for it 10-bit is good but once you start making any sacrifices you would probably be better off using 8-bit instead.

Thanks for the quick reply.

Don't worry about 3D settings, I'll sort it out when I'm back. I don't know how it compares to 1080p60, as I almost always use 1080p23, but I do know that I had to create a specific profile to get it to work on my HD7870, as my 1080p23 settings were too taxing.

Re 10bits, I did some testing and the difference is minimal. It's more for a "peace of mind" than a true PQ aspect. I don't want to use the OS HDR setting as it means I would be in HDR all the time, which I don't want. I want to use the nVidia dynamic switching mode as soon as the driver is fixed.

I need to try FSE with the 1080 Ti, I didn't use it (except when testing HDR) with the HD7870 as my JVC needed 15sec or more to sync, especially when using OSD, so I found the improvement in PQ wasn't worth the downside. However, my HD7870 had no frame drop for 2 hours+, while the 1080 Ti has frame drops every 3-5mn (with rendering times way below 30ms in 23p), so I might have to use FSE anyway.

I'll update here when I'm back and have had some time to play with the 1080 Ti more. I only had a couple of hours to test it before going on holiday. :)

[Asmodian]

After acquiring a new display (LG OLED 55" C7P) I changed my settings a lot. This OLED TV does not like NGU Sharp at all, even bluray 1080p looks artificial when using it, but NGU Anti-Alias is very natural and looks great. I also do not use the ED dithering options with it; I was unable to see any differences with ED dithering, ordered dithering is faster and offers identical quality on this display.

I was surprised by the similarity calibrating this OLED TV to my plasma TV from 2010. The both dither a lot to display shadows as well, they seem to be oddly similar technologies.

settings.bin (http://www.mediafire.com/file/2c118c36s25w554/settings.bin) updated.

[MariaX9]

Question because I do not understand it. I have at the reduce banding artifacts setting low/medium, should i tick "dont render frames when fade in/out is detected"?

[huhn]

well the LG OLED screens are more known for not dithering at all than doing heavy dithering.

http://i.rtings.com/images/reviews/c7/c7-banding-large.jpg
http://i.rtings.com/images/reviews/a1e/a1e-banding-large.jpg

[Asmodian]

Question because I do not understand it. I have at the reduce banding artifacts setting low/medium, should i tick "dont render frames when fade in/out is detected"?

I do tick "dont render frames when fade in/out is detected". If you are using easy settings for your GPU you do get 5 more medium debanded frames during the start of a fade with it not ticked but your GPU has to render 5 extra frames every time it detects a fade. This spike in load can cause a dropped frame or presentation glitch.

well the LG OLED screens are more known for not dithering at all than doing heavy dithering.

http://i.rtings.com/images/reviews/c7/c7-banding-large.jpg
http://i.rtings.com/images/reviews/a1e/a1e-banding-large.jpg

It definitely dithers to display shadows, like a plasma, you can see it quite easily.

Banding was a huge problem when setting it up but the banding does not mean it is not dithering, only that it is bad at it. When I set it to 6 bit in madVR the display's banding goes away pretty much entirely, even with default settings, and when set to 10 bit the banding is more common (I can see more of the steps). Banding test patterns all look very different at 6, 8, or 10 bit so I am sure the display is receiving different bit depths but it seems to have no idea how to go about displaying them without banding. Rounded looks much worse all the time, each step is visible and sharper, so it is not simply rounding anywhere. How do you get visible banding with a 16-bit gradient dithered to 10-bit on a 0-95 cd/m^2 display?!? :mad:

I had to tweak contrast and brightness a lot before finding settings that minimize the banding, which helped but still didn't eliminate it completely. I also have to be careful when watching real content and blaming banding on the display, some blurays I thought were fine turned out to have a lot of banding while others do not. If high debanding in madVR fixes it I blame the source but if it doesn't change it noticeably I blame the display. For my final calibration the contrast ratio was reported as 9528804659:1 by CalMAN, which does expose banding in the source.

All together this display does look very good, the black level and color saturation are both amazing, but banding is its major weakness. :(

[Manni]

Just a quick update re 3D, nVidia and MadVR.

The key to get this working seems to be to enable steroscopic 3D in the nVdia drivers while in 1080p, run the nVidia 3D Setup, and then set the desktop to whichever resolution/frame rate you actually use. My default is 4K@23p. MPC-BE/LAV/MadVR switch to 1080p3DFP to play the film automatically. They don't go back to 4K23 automatically, I have to leave MyMovies for that to happen. Sometimes they don't go back to 2D either, but in that case I only have to start playing a 2D movie.

The experimental custom rs tool Madshi made available recently has done wonders for my 2D bluray playback: I went from one frame repeat every 3-5 minutes to 1 frame drop every 2 hours or more, so very happy with that.

I still have to play a bit with the settings to find what works best with my new 1080 Ti, I really like the bluray to 4K upscaling using NGU Anti-Alias.

Overall super happy with the upgrade to the 1080 Ti from my HD7870.

Thanks again for all the help.

[Asmodian]

Thanks for the update on 3D!

I too have been very happy with NGU Anti-Alias. :)

[cyber201]

hi Asmodian, I'm waiting your settings for the new version of madvr....
Yeahhhh

Thanks
Bye

[Asmodian]

Finally updated for 92.2. A lot of fixes along with the amazing new custom resolution tool in display modes. :D

Also DX11 native DXVA2 decoding, no quality loss but still no IVTC or black bar detection.

I also rebuilt my settings, they are still the same settings but I reset to default with the new version and reconfigured from scratch. There were some people reporting crashes with old settings files (and even those looks to have been fixed in 92.2) and while I didn't have any issues I figured a fresh start was the safest option.

[aufkrawall]

Native DXVA2 decoding doesn't show the chroma blurring when DXVA2 chroma scaling is used at the same time.
Since DXVA2 scaling seems to equal some kind of bicubic on Nvidia, it's actually not too bad afterall and dimensions better than bilinear.
There don't seem to be any other quality drawbacks with recent madVR version on Windows 10 Creators Update, no banding problems.

New DX11 decoding API btw. isn't called DXVA anymore, afaik, but D3D11VA.

[madshi]

FWIW, I'm not sure if the custom modes are really limited to 8bit. This will depend on the driver. So it might be totally different with Nvidia, AMD and Intel.

[Manni]

FWIW, I'm not sure if the custom modes are really limited to 8bit. This will depend on the driver. So it might be totally different with Nvidia, AMD and Intel.

Definitely not limited to 8bits with nVidia. I have a custom mode running for 4K23p @ 12bits 4:4:4

[Asmodian]

Can you set the custom mode to be 12-bit in the Nvidia control panel? I can also get 12-bit RGB to work but I have to be in a 12-bit mode already before switching to it. I do not have the option to set the custom mode to 12-bit in the control panel.

Native DXVA2 decoding doesn't show the chroma blurring when DXVA2 chroma scaling is used at the same time.
Since DXVA2 scaling seems to equal some kind of bicubic on Nvidia, it's actually not too bad afterall and dimensions better than bilinear.
There don't seem to be any other quality drawbacks with recent madVR version on Windows 10 Creators Update, no banding problems.

New DX11 decoding API btw. isn't called DXVA anymore, afaik, but D3D11VA.

I tried to capture this information, is there anywhere you think I need to add something? Thanks. :)

FWIW, I'm not sure if the custom modes are really limited to 8bit. This will depend on the driver. So it might be totally different with Nvidia, AMD and Intel.

Good point, I will edit my entry to reflect my limited knowledge of the situation. :)

[huhn]

Native DXVA2 decoding doesn't show the chroma blurring when DXVA2 chroma scaling is used at the same time.
Since DXVA2 scaling seems to equal some kind of bicubic on Nvidia, it's actually not too bad afterall and dimensions better than bilinear.
There don't seem to be any other quality drawbacks with recent madVR version on Windows 10 Creators Update, no banding problems.

New DX11 decoding API btw. isn't called DXVA anymore, afaik, but D3D11VA.

doing a short test:

dxva: https://abload.de/img/dxvaoeob0.png
bilinear: https://abload.de/img/bilineary8of4.png
bicubic: https://abload.de/img/bicubicz8qci.png
they are all the same frame the shift comes from the scaler.
it is as bad as ever and bilinear is doing better.

[aufkrawall]

wth is this example even supposed to portray? It's far off any reality.
It's completely nonsense to think chroma would be blurrier than with bilinear.

software decoding + bilinear:
https://abload.de/thumb/softwarebilj8sr6.png (http://abload.de/image.php?img=softwarebilj8sr6.png)

DXVA2 native decoding + bilinear:
https://abload.de/thumb/dxva2bil5ms6q.png (http://abload.de/image.php?img=dxva2bil5ms6q.png)

DXVA2 native decoding + DXVA2 scaling:
https://abload.de/thumb/dxva2dxva4asou.png (http://abload.de/image.php?img=dxva2dxva4asou.png)

Oh, and bilinear downscaling also gives you one aliased mess for every target which is not exactly 1/4 or 1/16 etc. of source resolution btw...

[huhn]

so this BD is not a real source? yes chroma doesn't matter much for this image but that doesn't change what it is doing to the chroma channel.

what this is suppose to show?
obviously that there is something terrible wrong with DXVA chroma scaling.
just to make that clear i was only using DXVA chroma scaling nothing else in that image there is a reason i add the OSD.

and it is nonesense to think it is blurrier than bilinear? it clearly is in this case...

and about banding DXVA is still using low bitdeep and is adding banding you can check this with the BW.avi. where this will be notable is a different story but still it is still terrible at it.

[aufkrawall]

so this BD is not a real source?
yes chroma doesn't matter much for this image but that doesn't change what it is doing to the chroma channel.

obviously that there is something terrible wrong with DXVA chroma scaling.

and it is nonesense to think it is blurrier than bilinear? it clearly is in this case...

The image is grayscale only (at least to the eye), but actually even between bicubic and bilinear chroma scaling there is a huge difference in sharpness.
So what does it prove? Only that a soft chroma scaler softens the image here. There is already lots of aliasing in the source, and of course a soft scaler has a strong influence on that.

If there was a real problem with chroma blur & DXVA scaling, my cartoon example would look like blurry garbage. But actually, both chroma and luma are way sharper than with bilinear (it's 640x368 -> 2532x1440).

I have the suspicion that DXVA2 scaling uses different sharpness levels for different source resolutions. I tried a 720p non-cartoon sample a few days back and here DXVA2 scaling behaved mostly like madVR's softcubic60.
This is not bad quality, it may be just not to your likings.


and about banding DXVA is still using low bitdeep and is adding banding you can check this with the BW.avi. where this will be notable is a different story but still it is still terrible at it.
I don't see it:
https://abload.de/thumb/dxvaxps26.png (http://abload.de/image.php?img=dxvaxps26.png) https://abload.de/thumb/softwareq4s3o.png (http://abload.de/image.php?img=softwareq4s3o.png)

[huhn]

my screens are Cr rendered as Y.
it's an madVR feature to look up the chroma channels directly.
so if i want to know it DXVA got fixed i just use that.


but the real issue is that DXVA is known to ruin the chroma channel on nvidia cards and in my test nothing has changed so it is still bad in my book.

and now about your "10" bit image. first of all it is pretty much a solid color so a chroma scale is not adding a lot if at all banding here. second the source has banding you can see it in the 10 bit parts all 1024 steps and this get's removed by madVRs debanding making it a really bad test pattern for banding. that fact that the 8 bit is rounded is the nail in the coffin not a fair comparison at all.

[Manni]

Can you set the custom mode to be 12-bit in the Nvidia control panel? I can also get 12-bit RGB to work but I have to be in a 12-bit mode already before switching to it. I do not have the option to set the custom mode to 12-bit in the control panel.


I can absolutely select 12bits in the nVidia control panel when my 2160p23 custom mode is selected. It's the one made from the EDID values, as none of the others would work for me here.

Of course I have to select "use nVidia colour settings" not "use default color settings" for this to work, but that's the same with non-custom modes so you should already have this selected.

[Manni]

Correction: I was still under 385.28. I updated to 385.41 and I now have the same behaviour (can't select 12 bits from custom mode).
Reverting to 385.28

[aufkrawall]

my screens are Cr rendered as Y.
it's an madVR feature to look up the chroma channels directly.
so if i want to know it DXVA got fixed i just use that.

Nope, I'm sticking to real world tests (except of gradient images).
Especially since we are talking about DXVA scaling for low-end GPUs (it's called compromise).


but the real issue is that DXVA is known to ruin the chroma channel on nvidia cards and in my test nothing has changed so it is still bad in my book.

You haven't told me why the blur would be bad, nor why I do not encounter it with my cartoon example, where the blur by DXVA2 decoding is easily visible. Compare with softcubic?


and now about your "10" bit image. first of all it is pretty much a solid color so a chroma scale is not adding a lot if at all banding here. second the source has banding you can see it in the 10 bit parts all 1024 steps and this get's removed by madVRs debanding making it a really bad test pattern for banding. that fact that the 8 bit is rounded is the nail in the coffin not a fair comparison at all.
Provide a test sample please.

[nevcairiel]

but the real issue is that DXVA is known to ruin the chroma channel on nvidia cards and in my test nothing has changed so it is still bad in my book.

That isn't an inherent property of DXVA scaling or anything, its just an artifact if you try to access the 4:2:0 "untouched", because there is no 4:2:0 texture format in D3D9, but when you use DXVA to upscale chroma (or the entire image) anyway, then this doesn't really apply. Its not like the chroma is blurred first and then upscaled, thats not how it works. The method madVR uses to try to extract 4:2:0 subsampled chroma just causes blurring.

[aufkrawall]

That isn't an inherent property of DXVA scaling or anything, its just an artifact if you try to access the 4:2:0 "untouched", because there is no 4:2:0 texture format in D3D9, but when you use DXVA to upscale chroma (or the entire image) anyway, then this doesn't really apply. Its not like the chroma is blurred first and then upscaled, thats not how it works. The method madVR uses to try to extract 4:2:0 subsampled chroma just causes blurring.
That explains why DXVA2 chroma scaling "fixes" the blur with native DXVA2 decoding in my cartoon example. :thanks:
It's cleary visible that it totally disppears.

[Asmodian]

Correction: I was still under 385.28. I updated to 385.41 and I now have the same behaviour (can't select 12 bits from custom mode).
Reverting to 385.28

Ack! That doesn't seem like something Nvidia would change on purpose but it would also be a strange bug to randomly appear. :confused:

Thanks for the update.

[huhn]

That isn't an inherent property of DXVA scaling or anything, its just an artifact if you try to access the 4:2:0 "untouched", because there is no 4:2:0 texture format in D3D9, but when you use DXVA to upscale chroma (or the entire image) anyway, then this doesn't really apply. Its not like the chroma is blurred first and then upscaled, thats not how it works. The method madVR uses to try to extract 4:2:0 subsampled chroma just causes blurring.

well the image you see there is pure DXVA nothing elsed and the OSD isn't indicating that madVR is accessing the 4:2:0 image.
using DXVA native and no DXVa chroma is a whole different story.

Provide a test sample please.

take your screen deband it and the banding is gone.

if you want to see banding from DXVA image scaling take the BW.avi output it as 4:2:0 and see for yourself.

but the most important part i'm not even sure if DXVA scaling is any faster than lanczos 3 the way madVR is handling it.

[aufkrawall]

take your screen deband it and the banding is gone.

There is no banding for the 10 bit stripes without deband, and neither there is any for them when you convert the video to 8 bit.

[huhn]

"you should see 1024 vertical bar if your display supports 10 bit"

so there are not 1024 steps of banding in there (i can see them on my TV and they say they are there but...) but that's not banding because what again?

and if you really think 1024 steps dither to 8 bit is enough to hide them well go ahead and set madVR to a 10 bit internal processing.

[aufkrawall]

"you should see 1024 vertical bar if your display supports 10 bit"

so there are not 1024 steps of banding in there (i can see them on my TV and they say they are there but...) but that's not banding because what again?

Well, you initially mentioned a "BW.avi". Where can I find this sample if it's any better?
You said that DXVA scaling would show banding, so where do you see it?

The 1024 step video is totally enough to make EVR or Win 10 video app show banding, which is not there with madVR + DXVA.

[huhn]

there is clear 1024 steps of banding on this file. i can see the banding on your screenshoots. 16 bit dither to 8 bit is clearly better than 16 bit rounded to 10 bit.

the bw.avi is a file that allows you to see the used scaler.
EVR normal does a better job than madVR with DXVA scaling so that's important at least on my system: https://abload.de/img/bw04une.png
and it is bilinear right now (385.41) which is odd.

here you go: http://filehorst.de/d/cmdBtJyd
but not sure why you want it because it is not a real video.

[aufkrawall]

Can you tell me where do you spot banding here (deband off)?

Software decoding + bilinear:
https://abload.de/thumb/softwarebil2csbg.png (http://abload.de/image.php?img=softwarebil2csbg.png)

DXVA2 native + DXVA2 scaling:
https://abload.de/thumb/dxva2dxva2sc6zsb9.png (http://abload.de/image.php?img=dxva2dxva2sc6zsb9.png)

https://abload.de/thumb/softwarebilstatnxsam.png (http://abload.de/image.php?img=softwarebilstatnxsam.png) https://abload.de/thumb/dxva2dxva2scstatm2si7.png (http://abload.de/image.php?img=dxva2dxva2scstatm2si7.png)

The only difference I see is that bilinear is softer here, so the grain in the gradients is harder with dxva2. But I don't notice additional banding.

-> Still not noticing any "real world" problem.

[huhn]

there is no real banding in this example.
but you are downscaling a 10 bit source it's harder to add banding there.

nvidia 384.41:

bicubic: https://abload.de/img/bicubic6034rgk.png
DXVA: https://abload.de/img/dxva1yu99.png

the last thing i have to say about the DXVA nvidia scaler avoid it if possible...

[aufkrawall]

Well, it looks like softcubic. At least I don't see quality issues which would prevent me from using it over bilinear when using a lowend/integrated GPU.