madVR Options Explained

[foxyshadis]

maxkill, I think it's obvious that you're in the minority here, and that your opinion on what most people like might not be true. (For an extreme example, I'm not sure anyone wants to go back to the VHS era.) Please drop this pointless argument, no need to be defensive about what looks best to you, no one is telling you that you can't watch the way you want.

[v0lt]

@Asmodian
I do not see the pictures with the madvr working scheme on the first page of the topic. Can you return it?
Can you add an HDR-> SDR conversion to the scheme? I want to understand if this is done before or after resize.

[huhn]

you can check the OSD it should be done at the end like all color corrections.

[Asmodian]

I found a new host that doesn't watermark the images. :p

Like huhn I assume it is done at the end, right before the 3DLUT. I don't feel comfortable placing it on an image because we don't actually know exactly when it happens. If anyone has any evidence for when HDR->SDR happens I would be happy to take a look. :)

[Asmodian]

After some experimenting with HDR content and "ShowRenderSteps" I have determined that HDR tone mapping actually happens before scaling. This is nice for performance if you do upscaling but bad if you downscale. Tone mapping before scaling might be important to avoid clipping issues after scaling (edit: though that sounds backwards to me...).

I added a new flow chart for UHD HDR content. :)

[chros]

Nice, thanks!

[zerowalker]

I also wish you would not promote your settings here, they are so much in the wrong direction from what I think is good practice they almost seem like a parody.

They also seem like settings made by someone very new to quality video on PC and madVR. I see these overprocessed videos a lot and find it depressing. :(

If you like it fine, you can use your settings, but expect pushback if you promote them. I am just expressing my opinion too. I don't want doom9 promoting bad newb madVR settings as great after all. :p

I would like to hear your suggestion on settings.
I don't care much for processing as i don't want to alter the video,
but rather the up/downscaling, rendering etc.

I currently have:

Chroma upscaling: NGU - Standard, high quality
Image downscaling: SSIM - 2D - strength: 100% - activate anti-ringing filter [relaxed]
Image upscaling: NGU Standard - very high

Rendering - user D3D11 - use a frame for every VSync
CPU Queue 128
GPU Queue 24
All off on trade quality for performance

I basically want as high quality without messing up the video.
I chose these settings (especially the scaling) as i thought they were the best (meaning not the best for me, but best quality of the options available).
I couldn't use NGU on very high for chroma as it seems to use up my GPU sadly enough (GTX 1070).

The queues i am unsure about, from the explanation i guess CPU queue has a hand in subtitles,
and one kinda frustrating bug/issue i have is that when jumping back/forth in a video the subtitles sometimes don't render,
so i have to do another back/forth to re-trigger it.
It would be awesome if that could be solved,
but i don't think many run into it as probably jump around in videos a lot more than normal people do haha xd

Thanks:)

[Asmodian]

I think your settings are generally quite good.

Chroma upscaling: I prefer NGU AA, especially for chroma. Even when using NGU sharp I use NGU AA for chroma. I never set it higher than high for chroma, even with a 2080 Ti and in my SD profiles. Very high doesn't seem to improve real world chroma scaling, it may even harm it. So don't feel like you need a better GPU for that. :)
Image downscaling: I actually use SSIM - 1D 100% LL AR relaxed, 2D can be a bit much depending on the content and scale factor. 2D is great sometimes but I like 1D more for most 4K -> 1080p downscaling when I want to watch 4K content in a window. Linear light is important for downscaling.
Image upscaling: I like NGU AA very high, especially for animated or non-real life content but NGU AA v.s. NGU Standard is a fairly subtle difference with trade-offs either way. I don't like upscaling and downscaling for scaling factors below 1.5 for doubling or 3x for quadrupling. I also use Jinc AR for upscaling and Lanczos3 LL AR for downscaling after doubling/quadrupling.

There is no reason to set particularly large buffers, it simply spikes the IO and CPU/GPU for longer after a seek or on startup. As long as they are big enough to never get frame drops due to background tasks I set them as low as possible. The defaults are quite good and I use 8/16 for GPU/CPU buffers. I also use D3D11 with present for every vsync.

One setting you forgot to mention was the frames presented in advance. I found 3 to work best for me on Win10 and current Nvidia drivers (436.48), again as low as possible while never getting presentation glitches or frame drops. 8 seems too high.

Dithering is also very important. On my LG C7 and C9 I like ordered dithering but on my previous LCD screens I preferred ED1 or 2. I also much prefer disabling both colored noise and change dither for every frame. In my "Max Quality" settings.bin, linked at the end of the third post in this thread, I set ED1 for "slow" dithering but I actually use ordered dithering for both fast and slow myself. I just couldn't post a "max quality" settings file and not use one of the error diffusion dithering options. :o

[DMU]

Why is the rendering time significantly longer in OS HDR than in AMD HDR?
Screenshot. (https://drive.google.com/open?id=1R8GkNgra9ftQxm-9iWcXLhsRdIVdhF1G)

[chros]

Dithering is also very important. On my LG C7 and C9 I like ordered dithering but on my previous LCD screens I preferred ED1 or 2. I also much prefer disabling both colored noise and change dither for every frame.
That's interesting, can you explain why? Do you have any content to reliably test it with?
I used ED2 (both options were checked) for ~2 months now and sometimes noticed that something is odd, although I couldn't really tell why (maybe it was the source as well).
Now I switched back to Ordered dithering and I'll try to wait for a content to feel the above :)

[Cowboydude99]

Hey there
Can someone help me figure this out!

Why is my video concerting to a different resolution?

I don’t see any setting with that resolution...

The movie won’t really play. A few frames a minute..

[Asmodian]

Please attach images somewhere else and post a link, it takes a long time to get attachments approved here. :)

That's interesting, can you explain why? Do you have any content to reliably test it with?

I don't really. Testing dithering options at 8 bit or higher always feels a bit... placebo?

I have noticed colored dithering being on before, and been very happy to be able to turn it off, but I couldn't actually say how I noticed or why I cared. :o

[nsnhd]

Image upscaling: I like NGU AA very high, especially for animated or non-real life content but NGU AA v.s. NGU Standard is a fairly subtle difference with trade-offs either way. I don't like upscaling and downscaling for scaling factors below 1.5 for doubling or 3x for quadrupling. I also use Jinc AR for upscaling and Lanczos3 LL AR for downscaling after doubling/quadrupling.
With NGU Standard you can do quadrupling, but not with NGU AA, so quadrupling is better for upscaling 720p to 2160p ?
Any reason to choose Lanczos3 LL AR instead of SSIM - 1D 100% LL AR or Bicubic 150 LL AR after doubling/quadrupling since they cost the same performance I think ?
Thank you.

[Asmodian]

With NGU Standard you can do quadrupling, but not with NGU AA, so quadrupling is better for upscaling 720p to 2160p ?

Not in my opinion. NGU direct quadrupling is lower quality than NGU 2x+2x, but it is also faster. This is a complicated performance/quality tradeoff but if you can do NGU very high 2x followed by NGU high 2x it should be better than NGU very high direct 4x.

Any reason to choose Lanczos3 LL AR instead of SSIM - 1D 100% LL AR or Bicubic 150 LL AR after doubling/quadrupling since they cost the same performance I think ?
Thank you.

Lanczos is much easier than SSIM 1D, but it is actually a little harder than Bicubic 150. I find downscaling after quadrupling does not noticeably benefit from SSIM and lanczos is quite fast and offers mostly artifact free downscaling. Bicubic 150 is great when you need it really fast but it not superior to the spline family (lanczos/spline) in my experiance.

SSIM can even preserve NGU artifacts too well.. at least maybe in ways I would have trouble proving... but it offers no obvious improvements. :o

[Klaus1189]

Hey there
Can someone help me figure this out!

Why is my video concerting to a different resolution?

I don’t see any setting with that resolution...

The movie won’t really play. A few frames a minute..

Please attach images somewhere else and post a link, it takes a long time to get attachments approved here. :)

As Asmodian already said, upload also here to Google drive or whatever so we can view it.

[nsnhd]

Not in my opinion. NGU direct quadrupling is lower quality than NGU 2x+2x, but it is also faster. This is a complicated performance/quality tradeoff but if you can do NGU very high 2x followed by NGU high 2x it should be better than NGU very high direct 4x.
Except for NGU AA, I can't find the way to set NGU 2x+2x, can you ?
But in my eyes NGU Sharp very high 2x > Jinc AR looks better than NGU Sharp very high direct 4x <SSIM 1D. You're right, thanks.

[huhn]

NGU AA is always x2 and you can use it to do x2 two times.

[nsnhd]

NGU AA is always x2 and you can use it to do x2 two times.
Yeah, but I can't do it for other NGU's

[Asmodian]

Yeah, but I can't do it for other NGU's

Yes you can, simply set luma quadrupling to 'double again'
https://i1222.photobucket.com/albums/dd496/asmodian3/madVR%20-%20double%20again.png

[nsnhd]

Yes you can, simply set luma quadrupling to 'double again'

Thanks, upscaling 720p to 2160p works for NGU very high-NGU high with gtx 1060 6gb, though I have to lower the chroma of NGU to Normal (Bicubic60 AR)

[cbnbmore]

I am trying to settle in on some solid settings now. What Driver do most of you use for the RTX cards now?

[Asmodian]

I like 441.20 on my 2080 Ti, at least for my uses. The latest at the time of posting.

[Bartok]

hi Asmodian,have you got some advices to take benfit of my pana FZ-800 for hdr settings: tone map or passtrough ? Calibration settings best choice ? 8,10,12 bits usage ? By the way i'm using you high profile wirh rtx 2080 , works fine :)
thanks a lot !

[ng4ever]

436.30 imo


I left this forum because I got the best settings for madVR from a certain forum poster for my RTX 2060 Super ! Thanks again.

Can now do 360, 720, and 1080p upscaled easily.

[petoulachi]

Hello Asmodian,

first of all thanks for sharing your settings, it's a wonderful base to start to do advanced filters with MadVR.
I have one question regarding your devices, in the calibration tab. With your C7, You are using 3DLut, and havce setup BT709, but also SMPTE C,EBU/PAL and BT2020 (with C9 I guess you have set the LUT directly onto the TV). Could I ask you what software did you use to do them, and when are there usefull ?
I'm using DisplayCal/Argyllcms and by default it generate only BT709, and I'm curious to know if there are some points that could be better on my configuration.

Also, why don't you use 3DLut for hdr ?

You also switch to 10bits/8bits depending on the HDR or not, but not with your C9.


Thanks for your precious input :)

[zerowalker]

About the CPU/GPU queue size, is there some way to determine if there's framedrops (i mean if it occurs always it's easy to spot, but i think i have seen smooth motion not working properly at times, but just jumping forth/back or something to reset it fixes it, and i think it might fix itself, it's not that obvious so not really sure if i am just mistaken though).

And also how does CPU/GPU differ there, as the CPU is the one it's limited to i am guessing they act differently and the CPU is the real buffer and the GPU is a some backup somehow?

Also don't recall if i asked this before, probably did, but is the "use a separate device..." preferred, or is it just overhead for no real reason?

[Asmodian]

Are you using some form of VRR? G-sync/Freesync? Smooth motion does not work well with VRR, disable VRR for madVR and/or your player.

The CPU buffer is the in system memory buffer. It holds frames before they are uploaded to the GPU, after any CPU processing steps (IVTC, etc.). This is the "before madVR" buffer.

The GPU buffer is for after madVR but before present, this is how many completed frames are stored on the GPU. The "after madVR" buffer. This one has more impact in my experiance, the CPU buffer only benefits from a few extra frames. Too big is as bad as too small in my experiance, the default 8 is good. 6-12, above that has never helped on my systems.

Use a separate device does not seem to help with Windows 10. It was mostly useful in the early days of hardware decoding on the GPU, newer OS and drivers seem to have fixed any issues that using a separate device might help. I leave it disabled today as unnecessary overhead but it did not seem to hurt anything during testing.

[maxkill]

Getting stutters from 60fps video to 23hz playback. Normally I play 24hz videos etc no problem, but get stutters when playing the 60fps movie. (My screen is set to 23hz with max 30)

I get no performance spikes to 100% and I tried enabling the hack direct3d with no success.

Any ideas what settings to make the 60fps video smooth on under 30hz?

[huhn]

well you can't show a 60 fps file smooth at 23p that's simple not doable.
at 30 it could work out if the source is true 60 with no 30 fps parts or it will be a disaster.

not sure if madVR has proper decimation it's nothing user usually need.

[maxkill]

well you can't show a 60 fps file smooth at 23p that's simple not doable.
at 30 it could work out if the source is true 60 with no 30 fps parts or it will be a disaster.

not sure if madVR has proper decimation it's nothing user usually need.
It's at
59.940 fps
to be exact

What settings to make it smooth on 30hz<? Or impossible?

[huhn]

just use 60 hz.

[Asmodian]

Yes, madVR's smooth motion is not designed to handle frame rates above the refresh rate. They must be the same at least, but really the refresh rate should be more than double the frame rate before smooth motion works well.

[huhn]

well i personal think it is not useful for that but it has extra code for this situation and when SM was new i talk about it well i even started a bug report before the talk started no one cared so it was left in like this and it does what you see intentional.

so technically it was mode for this. i simply don't know why someone would want that result.

[Asmodian]

When first designed it was thought about but I think it has never been optimized for it to actually give a good experiance. Since no one cared it was dropped from the design. ;)

It seems to 'randomly' drop frames instead of maintaining smooth but blurry blending. I think it cannot blend more than two frames so it seems choppy when some frames get dropped.

[huhn]

AFAIK it is completely finished.
i saw it blending 5 frames.

[DMU]

Any ideas why the rendering time in the OS HDR mode is much longer (https://drive.google.com/open?id=1R8GkNgra9ftQxm-9iWcXLhsRdIVdhF1G) than in the AMD HDR mode?

[chros]

Display is at 60Hz, do you use smoth motion? Maybe it switches back to copyback?

[DMU]

Display is at 60Hz, do you use smoth motion


In this case, no.

Maybe it switches back to copyback?
All settings are the same except for the Windows HDR switch.

[maxkill]

Well, it seems my receiver is capped at 30hz so I'm shit out of luck then.

[huhn]

that would break the spec try it at 1080p

[DMU]

Has anyone dealt in detail with the option "scale chroma separately, if it saves performance"? Can you comment on this (https://forum.doom9.org/showpost.php?p=1896839&postcount=58438) post?

[zerowalker]

Are you using some form of VRR? G-sync/Freesync? Smooth motion does not work well with VRR, disable VRR for madVR and/or your player.

The CPU buffer is the in system memory buffer. It holds frames before they are uploaded to the GPU, after any CPU processing steps (IVTC, etc.). This is the "before madVR" buffer.

The GPU buffer is for after madVR but before present, this is how many completed frames are stored on the GPU. The "after madVR" buffer. This one has more impact in my experiance, the CPU buffer only benefits from a few extra frames. Too big is as bad as too small in my experiance, the default 8 is good. 6-12, above that has never helped on my systems.

Use a separate device does not seem to help with Windows 10. It was mostly useful in the early days of hardware decoding on the GPU, newer OS and drivers seem to have fixed any issues that using a separate device might help. I leave it disabled today as unnecessary overhead but it did not seem to hurt anything during testing.


Is there a way to benchmark it to see what works best?
Cause it's very hard to just See so to speak, even if i set it all to the lowest possible it still works most of the time from what i can tell.
So it would be neat if i could check for framedrops or something in a log perhaps, not sure how to stresstest it though,
i am guessing as heavy video decoding as possible.

[Asmodian]

If it seems to work fine most of the time on the lowest settings I would simply use the defaults and be happy.

Don't worry too much about buffer sizes, they are not that critical and the defaults are chosen well.

[Asmodian]

Has anyone dealt in detail with the option "scale chroma separately, if it saves performance"? Can you comment on this (https://forum.doom9.org/showpost.php?p=1896839&postcount=58438) post?

When we were testing it when it came out it was decided that for real life content it looked better disabled, and in cases where you always had to scale the chroma (e.g. 1080p on a 1440 display). Still, it is more a case of tradeoffs rather than obviously better one way.

Blurring chroma differently than luma can cause odd artifacts, so perhaps disabled is safer. That said avoiding scaling entirely is sharper. For 4K video on a 1080p display I can see leaving it on, but test with something besides text on a solid background too.

I don't think text on a black background is a great chroma test pattern for this feature, it is easy to notice subsampling but it is not very representative of real life content.

[DMU]

For 4K video on a 1080p display I can see leaving it on, but test with something besides text on a solid background too.

I don't think text on a black background is a great chroma test pattern for this feature, it is easy to notice subsampling but it is not very representative of real life content.
Can you recommend specific real life content for the test?

[huhn]

mostly video games and mostly if you upscale and downscale at the same time.

they are the reason why i think NGU chroma should always be used with NGU image doubling.

your test should be repeated with a simpler chroma scaler but i'm pretty sure it will still loose with no questions. was this test file correctly mastered with the correct chroma position?

[DMU]

When we were testing it when it came out it was decided that for real life content it looked better disabled, and in cases where you always had to scale the chroma (e.g. 1080p on a 1440 display).
In my practice, I have not seen the operation of this option when upscaling an image, only when downscaling (for example, 2160p -> 1080p)

was this test file correctly mastered with the correct chroma position?
Not sure. I received the video for the test in 2 different ways, but in both cases the results were the same:
1 way
- from the source .png image was received .jpg image in 4:2:0 format using the paint.net editor;
- using ffmpeg a video was created for the test with the default settings.
2 way
- using ffmpeg a video was created for the test with the yuv420p setting from the source .png image.

[ikarad]

-target peak nits: [200] Set the target display's peak white brightness. Above this value is used as the highlights. A lower setting increases the brightness of mid range values.
What is the best value to use for my display? Is it 200 as you described or another value?

My display is https://www.tftcentral.co.uk/reviews/acer_xb270hu.htm

Thanks.

[Asmodian]

200 is simply the default. This is a good starting point but adjust it to your preference. There isn't really a "correct" way to handle HDR -> SDR, you have to decide what looks best to you and different HDR mastering might change your preference for this value.

[Audionut]

I was under the impression that target peak nits should be set to the measured white level of the display. There are other ways to adjust mid range brightness, right?

There may not be a "correct" way to do HDR -> SDR, but madVR is doing HDR -> SDR. Telling madVR that the target peak nits is going to be 200 when the display is only 100, or 500, or whatever, doesn't seem appropriate.