madVR - high quality video renderer (GPU assisted)

[seiyafan]

For 720 to 2160, which one is usually higher in quality, double then upscale or quadruple then downscale?

[makakam]

I mostly watch 1080p sources on my benq w1070 which is projecting onto a 92" screen. My watching distance is about 2,30m and I appreciate what NGU does with chroma, it really makes a difference. I own a gtx960 and set NGU to high. Average rendering time is about 37ms, it's pretty high but as long as I don't get dropped frames I'm cool with it.

[Anima123]

For 720 to 2160, which one is usually higher in quality, double then upscale or quadruple then downscale?

Madshi has mentioned earlier, NGU will support quadruple in one step instead of iterating as it do now.

[kasper93]

Have you noticed high CPU usage when using NGU? https://i.imgbox.com/CNYlkAXK.png I have i7 5820K so it is ~88% of thread power so not much left here. Correct me, but most of workload should be GPU so is this some busy waiting? Or some memory copy back and forth? Probably not much can be done here but I figured I will ask :)

[huhn]

IVTC black boarder detection is running on the CPU are you using any of these?

[ryrynz]

Nice to hear from someone that understands electronics.

There's been plenty of people who have weighed in with similar information over the previous pages, myself one of them. I wish people would understand how a thread works, getting a bit of this recently..

[Oguignant]

Does anyone recommend me best settings for 2k to 4k upscaling of REMUX mkv files?

[Georgel]

This one thread for everything is boring to read, that's why there are some aggressive comments. It's difficult to find useful information/question between repetitive old questions from new/less advanced people or specific content from very advanced users.

I find this harder to read for everyone including Madshi, but he still prefers this so bad luck for guys like me.
I think I made a made a post with interesting (I think) stats which can allows users to verify if normal behavior or not. Already 10 pages away. Won't do this again. :D

I was also curious in this information!

I feel sorry if I'm not adding anything useful.....

I have exactly that setup GTX 1080 + madVR + mpcHC/mpcBE + SVP4 and is working flawlessly :D
Both on 32&64 bits setups btw.

Well, it's good to hear that for someone it actually works well. Mine crashes at random times and especially on seeking the file, but the crashes are still there after uninstalling it, only DDU - ing the driver then reinstalling fixes it. Then again, I have a feeling that being a software dev, I might have a different set of libraries installed and such that could cause this. Impossible to track what exactly does it, I tried to study the crash logs but nothing very relevant came out of it.

There's been plenty of people who have weighed in with similar information over the previous pages, myself one of them. I wish people would understand how a thread works, getting a bit of this recently..


Sorry for this...

Coil whine is something really scary when you hear it and I would want to prolong the useful life of anything I have as much as I can. I did the same for my DAPs even though I sell every one of them after an year and a half or so, it's just the spirit to ind every thing I own.



I feel like I am to blame for opening my mouth here..............

[har3inger]

Coil whine is completely harmless to GPUs. Sometimes, once your GPU ages enough, the coil whine will go away by itself. I wouldn't worry about it, although I understand how annoying it is to deal with.

In any case, if coil whine exists, it will generally happen under a good number of other high load situations. It's not really something that's fixable with software changes. You could try for an RMA?

[Asmodian]

Madshi has mentioned earlier, NGU will support quadruple in one step instead of iterating as it do now.

At the same time Madshi mentioned it will triple directly too, which would be ideal for 720p on 2160p. :)

[austonrush]

Some screenshots would be nice. With SR4, if you don't mind, so I can see if SR really handles this better than NGU.

But I do agree, if you have sources with high compression artifacts, NGU will try to interpret the artifacts as important image detail. The current NGU version really needs high quality sources to shine.


Here are 4 samples. I've found adding grain to SD content really helps remove the proccessed look. I'm using level 2 in these images.

I'm also NOT using smoothen edges for SD content it is not required. It is required for HD...

Which one looks best to you? I should mention I'm getting a much better render performance with ngu low versus nnedi3 128 with SR3

:) thanks!

http://imgur.com/a/8MY6H

[ryrynz]

Here are 4 samples.


@640x352 in JPEG format..

[burfadel]

madVR v0.90.20 released

http://madshi.net/madVR.zip

* added workaround for weird crash with latest AMD beta drivers

Is that workaround still present? madVR was working fine on my old R9-280X, but for the RX 480 it doesn't like DirectX 11 presentation mode etc. This is running latest insider build plus Crimson 16.11.4. I'm wondering whether the 'workaround' is actually causing issues with the latest card series on the latest insider Windows if the underlying issue has been rectified.

[huhn]

try an older version that doesn't have the workaround.

[RyuzakiL]

Does Madshi had his own recommended settings for MadVR? Because I know there are others like me who doesn't want to be a scientist just to try all the settings and combinations possible for MadVR, and would just like to have an optimized, author/creator recommended settings.

If anyone of you happens to spot Madshi's own recommended settings, then I'm all ears and eyes. Guess I'm just getting tired of trying all the possibilities and just want to spend that testing time into quality time with my collection of Anime's and Movies.

[pose]

RyuzakiL, read Warner306's guide, it has a section with solid recommended settings for low to high end video cards.

[Uoppi]

Does Madshi had his own recommended settings for MadVR? Because I know there are others like me who doesn't want to be a scientist just to try all the settings and combinations possible for MadVR, and would just like to have an optimized, author/creator recommended settings.

If anyone of you happens to spot Madshi's own recommended settings, then I'm all ears and eyes. Guess I'm just getting tired of trying all the possibilities and just want to spend that testing time into quality time with my collection of Anime's and Movies.

I know the feeling; I have simply too little time for any meaningful zooming-in-and-comparing-screenshots. But it's hard to give recommendations when a lot is up to preference.

Here are some reasonable starting points partly based on recent posts by madshi and others, according to which I've set up my madvr:

1. Use NGU low/med/high (depending on your system's performance) for luma doubling. Choma doubling not as "necessary".
2. For luma upscaler, pick Jinc AR if you have the horsepower or Catmull-Rom 75 AR for a less taxing option.
3. Set chroma upscaling to NGU (I can't see a difference in normal viewing between low/med/high) or Catmull-Rom 75 AR.
4. Downscaling: Bicubic 150 AR. When downscaling straight without upscaling/doubling first (i.e. native 4K to 1080p screen), you might want to try SSIM 1D or 2D.
4. For dithering, Error Diffusion 2 or ordered dithering (I can see no difference between these in normal viewing)
5. For SD content, use option "add grain", for HD (I'd say 720p and up) use "soften edges". Both either at strength 1 or 2.
6. If you have the horsepower and want some (sometimes placebo) improvement, feel free to crank up the settings but watch the temps!

[Ver Greeneyes]

2. For luma upscaler, pick Jinc AR if you have the horsepower
Just remember that this will also be the upscaler used for chroma doubling if you're only doubling luma, so if you turned off chroma doubling for performance reasons you might want to pick a less demanding image upscaler to go along with it.

[ajp_anton]

Since I pretty much always upscale, I thought it always said ">" for the algorithm in the OSD. So now I just realized that > is upscaling and < is downscaling. Isn't that kind of backwards? A > looks like it becomes smaller (because we read left to right) and < becomes bigger.

[Backflash]

So after messing around for a week this is the result on 1440p(I miss my 1080p monitor for lowres content purposes).
r9 390 nitro/i72600 win10 x64

http://imgur.com/efwtjWw
640x480 839kbsp

debandingl med/high no anti ringing
sharpen edges 1.0, thin edges 1.0, anti bloat 100, anti-ringing
crhoma NGU med, superes 1
SSIM 1D 100/AR/LL/AB100
Doubling luma always x2 super sampling NGU med
image upscaling Lanczos 3/AR/SL
image refiniment SR2 AR(no LL)

any advice on improvement?
no NNDI3 though I hate toasting my hardware

[pose]

So after messing around for a week this is the result on 1440p(I miss my 1080p monitor for lowres content purposes).
r9 390 nitro/i72600 win10 x64

http://imgur.com/efwtjWw
640x480 839kbsp

debandingl med/high no anti ringing
sharpen edges 1.0, thin edges 1.0, anti bloat 100, anti-ringing
crhoma NGU med, superes 1
SSIM 1D 100/AR/LL/AB100
Doubling luma always x2 super sampling NGU med
image upscaling Lanczos 3/AR/SL
image refiniment SR2 AR(no LL)

any advice on improvement?
no NNDI3 though I hate toasting my hardware
Remove all image enhancements and upscaling refinements except soften edges and add grain (i prefer both on 1). Ditch always - 2x supersampling. I really like smoothmotion, you might want to give it a try. Dithering: Error Diffusion 2.

[aufkrawall]

Why ED2 and not just ordered dithering?

[GCRaistlin]

list all display modes madVR may switch to: 1080p24, 1080p50, 1080p60


http://mir.cr/SX2EUCS8
ReClock reports 23.976 fps. madVR OSD reports 29.970 fps, deinterlacing off. According to the activated profile the following is true: 25<=deintFps<31.

[ ] disable automatic source type detection
madVR switches the display to 60 Hz.
[x] disable automatic source type detection (.) film mode
madVR switches the display to 24 Hz.


Why does forcing to film mode change things? Which FPS is correct - madVR's or ReClock's?

http://mir.cr/0TO7HKQ7
[x] PAL SpeedDown (in ReClock), [x] treat 25p movies as 24p (in madVR).
madVR switches the display to 24 Hz. According to the activated profile the following is true: deintFps=25.

Why isn't deintFps equal to 24?

[Backflash]

Remove all image enhancements and upscaling refinements except soften edges and add grain (i prefer both on 1). Ditch always - 2x supersampling. I really like smoothmotion, you might want to give it a try. Dithering: Error Diffusion 2.

do not like soften edges on anime, grain is good yes
smooth motion is kind of useless on 144hz monitor
put doubling to always when needed- yeah slightly less artifacts

still think sr2 AR looks good on solid colors, dunno maybe it's my monitor and eyes combination creates this placebo
thin and sharpen edges is meh, agree
error diffusion 2 is always on

[Uoppi]

I personally set the dithering option last. I choose Error Diffusion 2 over Ordered Dithering if I've got performance to spare.

That said, at my regular 8 bit output, I can discern no difference whatsoever between them from my normal viewing spot. I think I had to lower the bits to 6 or so before I saw the difference with my nose glued to the screen. So... ED2 for me is pure placebo but I don't mind as it makes me feel like the quality is better (same thing with my audiophile hobby). :D

Hmm... ED2 is by and large regarded as "better", right? :o

[Asmodian]

Hmm... ED2 is by and large regarded as "better", right? :o

I like ED1 more with my current monitor but I liked ED2 more with my previous one. Not that I feel too strongly about it, they are both good and very similar. ;)

[theDongerr]

Seems to be plenty of knowledgeable people on this thread so I'll take a chance someone here can help me. I have a JVC-RS600 projector and HTPC and recently sold my Oppo blu-ray player thinking I can simply rip my BRs to the HTPC and use jRiver/MadVr to scale 1080p to 2160, where my projector will then take it the rest of the way to an e-shifted 4k picture.

Even with simple doubling of sbrx-100 for both luma and chroma, image enhancements SR-3. I get a great picture, and I would be fine if it wasn't for just one thing:

My eyes are pretty sensitive to motion judder. Mind you, my projector does 23.976 fps and I am able to playback my content at the proper FPS (i believe it is correct anyway) with 0% clock deviation from player to display. However, my white-whale exists as a video signal that looks like film with the motion smoothness of 60hz/120hz but without that fake-plasticky-SOE. My projector, even at the lowest CMD (clear motion-drive) setting, looks like a mexican soap opera. Even with an Oppo 103D blu-ray player, I still experience the same (if not more) judder than when playing movies from my HTPC -- therefore I sold it.

I've been told a professional calibration could possibly reduce the judder/motionblur I experience and as a result I have a guy coming out in January. In the meantime, I intend to keep tinkering with madVr and NGU. I only have a Radaeon R9 280 (3GB OC X-fire something) and it does 1080p ->2160p via sbxr-100 (luma and chrome) with SR-3 without any problems. My next step is going to be to audition a gtx 1080 and see if that helps any.

Has anyone any advice for smoothing out the motion a little without trashing my film-quality image?

Thank in advance!

[anta777]

I like ED1 more with my current monitor but I liked ED2 more with my previous one. Not that I feel too strongly about it, they are both good and very similar. ;)

Error Diffusion – option 1- Windows 7 and new, DirectX 11 and monitor support 8 bit and high.
Error Diffusion – option 2 — all ident option 1, only monitor support 6 bit.

[pose]

Error Diffusion – option 1- Windows 7 and new, DirectX 11 and monitor support 8 bit and high.
Error Diffusion – option 2 — all ident option 1, only monitor support 6 bit.
Source?

[Crimson Wolf]

dithering/algorithm:-


None: Do not choose this unless you have really crappy GPU. Choose this only for performance reasons because the quality is crap.
Random dithering: Also do not choose this unless you have really crappy GPU. Choose this only for performance reasons.
Ordered dithering: Choose this if your GPU is not crap but do not support DirectX 11, or if you still happened to use Windows XP.
Error Diffusion – option 1: Choose this if you use Windows 7 or later, have a DirectX 11 discrete GPU with decent horsepower, and preferably use a monitor with 8-bit display bitdepth or higher.
Error Diffusion – option 2: Choose this if you use Windows 7 or later, have a DirectX 11 discrete GPU with decent horsepower, and use a monitor with 6-bit display bitdepth.



dithering/options:-


Use colored noise: Best used with the two error diffusion options above. Do not enable in conjunction with the first three dithering algorithms.
change dither with every frame: Enable this option


https://imouto.my/tutorials/madvr/

I don't use every suggested setting from this guide or know if they are accurate or not but on my wide gamut display, ED1 looks better. On my gaming TN monitor, I can not tell the difference so I just use ordered on that. YMMV.

And to people asking about the "best" settings... If you would like someone to suggest settings to you, please provide your PC specs, monitor/TV model#, etc., and if you can calibrate your display at the very minimum. In the end, trust in your eyes. If you can't tell the difference, does it matter?

@madshi I was testing the HDR functionality of madvr using some samples from the demo-uhd site (life of pi) and tried the "film & tv" app from the microsoft store and it recognizes the hdr video and does its own thing. I was wondering what it was doing since I couldn't set madvr to display the same way?

[huhn]

is there an notable performance difference between random and ordered dithering?
and even if your GPU is crap you should never choice none...

[theDongerr]

Clock deviation -0.00453%
Rendering 32.48ms
chroma upscaling: NGU Low
image downscaling: Bicubic 150 AR
image doubling: NGU low
chroma doubling: NGU low
image upscaling: Catmull-Rom AR
refinements: none

Scaling 1080p24 -> 2160p24

To my eye, things look good. I'm not sure what the significance of the rendering time is, but so far nothing is obviously broken and picture looks good.

ALSO, these random questions to post are b.s.

[anta777]

Source?

http://rufootballtv.org/manuals/madvr/

[thighhighs]

http://rufootballtv.org/manuals/madvr/This guide looks like translated version https://imouto.my/tutorials/madvr/. But with small changes or just outdated.

[Growdelan]

EVGA GTX 660ti

1. ON MOUSE OUT

Artifact removal:
- reduce banding artifacts: on
- default debanding strength: medium
- strenght during fade in/out: high

scaling algorithms:
- Chroma upscaling: super-xbr 100 AR
- image upscaling: Lanczos3 AR
- upsacling refinement: SR 3 AR / refine the image only once after upscaling is complete

smooth motion: on
dithering: ED 1 / use colored noise / change dither for every frame
trade quality for performance: only check "don't rerender frames when fade in/out is detected"

2. ON MOUSE OVER

Artifact removal:
- reduce banding artifacts: on
- default debanding strength: medium
- strenght during fade in/out: high

scaling algorithms:
- Chroma upscaling: NGU-low
- image upscaling: Catmull-Rom AR
- upsacling refinement: soften edges 1 and add grain 1 - on

smooth motion: on
dithering: ED 1 / use colored noise / change dither for every frame
trade quality for performance: only check "don't rerender frames when fade in/out is detected"

Train to Busan 1280x694 -> 1680x1050
- super-xbr 100 AR: 12ms
- NGU-low: 15ms

http://screenshotcomparison.com/comparison/191729

The Treatment 1280x536 -> 1680x1050
- super-xbr 100 AR: 9ms
- NGU-low: 12ms

http://screenshotcomparison.com/comparison/191816

For me, super-xbr is sharper.

[huhn]

this is mostly comparing lanczos 3 vs catmull-rom

[Crimson Wolf]

@Growdelan

If you are only using NGU on chroma, I wouldn't bother using it. NGU amazing for the luma doubling.

Try these settings:

chroma: superxbr-125 AR only no SR
image downscaling: catmull AR or bicubic 150 (same thing)
doubling: luma only NGU-low if upscaling is needed
image upscaling: lanzco 3-tap AR
upscaling refinement: NOTHING enabled except soften edges 1, NO SR


I'm curious if the GTX660ti can handle these settings for that particular video. If it can, does it look any better?

[austonrush]

@640x352 in JPEG format..

I clicked F5 in fullscreen mode within mpc-hc and that is what it gave me... How can I get a better quality screen shot from MPC-HC? thanks

[huhn]

file -> save image... change save as type to PNG.

or ALT + i and change save as type to PNG.

[Growdelan]

@Growdelan

If you are only using NGU on chroma, I wouldn't bother using it. NGU amazing for the luma doubling.

Try these settings:

chroma: superxbr-125 AR only no SR
image downscaling: catmull AR or bicubic 150 (same thing)
doubling: luma only NGU-low if upscaling is needed
image upscaling: lanzco 3-tap AR
upscaling refinement: NOTHING enabled except soften edges 1, NO SR


I'm curious if the GTX660ti can handle these settings for that particular video. If it can, does it look any better?

Yep, can handle.

My old settings chroma: super-xbr 100 AR, upscaling image: lanczos3 without doubling vs your suggestion

http://screenshotcomparison.com/comparison/191825

first setting, still look sharp

[AngelGraves13]

Ok, so after some testing, I've settled on NGU-low for Chroma and Doubling (always if needed) and Soften Edges 1 and Add Grain 4. The grain really helps hide compression artifacts.

I've disabled SR and Remove Ringing Artifacts as it seems to degrade the image with NGU.

[Asmodian]

image downscaling: catmull AR or bicubic 150 (same thing)

catmull AR is not the same as bicubic 150, catmull is bicubic 50.

[Crimson Wolf]

catmull AR is not the same as bicubic 150, catmull is bicubic 50.

Ah, I'm sorry. I've switched to using SSIM for downscaling a while back. I must have got that part wrong.

@Growdelan

I just noticed your display is 1680x1050, so not a 1.5x scaling from original 720 video. That along with soften edges might be reason NGU is looking less sharp. On my display, when I do 720p > 1080p scaling, no matter which image upscaling and upscaling refinement settings I try, it would look worse than luma double and downscaling with SSIM, so I switched.

[Growdelan]

@Growdelan

I just noticed your display is 1680x1050, so not a 1.5x scaling from original 720 video. That along with soften edges might be reason NGU is looking less sharp. On my display, when I do 720p > 1080p scaling, no matter which image upscaling and upscaling refinement settings I try, it would look worse than luma double and downscaling with SSIM, so I switched.

I tried on the TV, but the result the same :(

http://screenshotcomparison.com/comparison/191841

[Backflash]

I tried on the TV, but the result the same :(

http://screenshotcomparison.com/comparison/191841

I'm pretty sure that's what Bicubic 150 does.
Try SSIM 1D 100% AR/LL(optional)/AB 50 or 100% depending on your taste. I think 660 can handle it.

[huhn]

super XBR is sharper in this case but NGU has less ringing.

[Growdelan]

I'm pretty sure that's what Bicubic 150 does.
Try SSIM 1D 100% AR/LL(optional)/AB 50 or 100% depending on your taste. I think 660 can handle it.

SSIM 1D100% + AR, LL, AB 100% - did not help, lips and eyes sharper on super-xbr + lanczos3.

http://screenshotcomparison.com/comparison/191909

[aufkrawall]

I never understood why there is AB for SSIM.

[Uoppi]

In the end, trust in your eyes. If you can't tell the difference, does it matter?


Maybe it shoudn't but...

Looking at various zoomed-in screenshot comparisons, it's more or less easy to see some algorithms are simply better (in terms of artefacts etc.)

Even if you couldn't see those differences at all under normal viewing conditions (like I don't think I can for chroma, for example), just knowing you're using the "worst" settings may degrade the experience. If your GPU can handle it, why not use a decent algorithm "just in case"?

Because placebo is a very real and powerful phenomenon*, I see no reason why using "recommended best" settings wouldn't be a valid and acceptable way to proceed.

(* I've personally heard noteworthy differences between regular and high-end audio cables - only to learn the tester actually did not switch any cables at all... :D)

[hu1kamania]

Enjoying NGU for 720p to 1080p, great job Madshi!

I'm sticking with NNEDI3 for lower quality sources. NGU seems to bring out compression artifacts with it's extra detail.