madVR - high quality video renderer (GPU assisted)

[sauma144]

teaser:

Two new algorithms coming in the next build, probably out tomorrow:

blocking (http://madshi.net/deblockA.png) -> cleaned (http://madshi.net/deblockB.png)
mosquito (http://madshi.net/mosquitoA.png) -> cleaned (http://madshi.net/mosquitoB.png)

More deblock examples, please. :)

[el Filou]

:eek: those blocking and noise filters may make MPEG2 SD cable watchable again (damn you, cable provider! :mad:), this is fantastic.

[Asmodian]

aliased sources
what does that mean?

http://i1222.photobucket.com/albums/dd496/asmodian3/pA7uy_zpsah4zfaix.png

[70MM]

With "crop black bars" active, madVR actually cuts the black bar pixels away, so madVR has less pixels to process. With "crop black bars" unchecked, madVR is aware of which pixels belong to the black bar and which don't, but the pixels are not cut away, so they're still being processed, they might just be zoomed away, eventually.

.
madshi so does this mean if we tick the box and zoom the black bars away, the image would be better again since there is no processing going on of the black bars as well?

[dvd1]

http://i1222.photobucket.com/albums/dd496/asmodian3/pA7uy_zpsah4zfaix.png
thanks for the example, but you can not make an example with pictures of a video?
thank you very much

[huhn]

What temperatures should be of concern?

From running monitoring tools default GPU temp is 47* C and max its running at appears to be 69* C. From what I recall from the past 90-100+* C might be of concern.

CPU cores are reporting temperatures from 69*C to 88*C.


I'm not familiar with what a "tile" size is. They are small little lines (don't go very far across the screen) that seem to be randomly spread in their location.

88 sounds high but it is the cpu. gpu looks ok but that doesn't mean it is not defect.

one thing you could try is running a torture benchmark and see if you get the same artifacts to exclude a problem with madVR.


You're saying they're not 4:3, but maybe 3.9:3 or something like that? Must be the fault of the studio doing the encoding then.


dvd/broadcast is/was created with a "overscan" part: https://en.wikipedia.org/wiki/Safe_area_(television)

there can be anything in there cameras uneven blackbars random noise what ever not a mastering error just a bad spec.

madshi so does this mean if we tick the box and zoom the black bars away, the image would be better again since there is no processing going on of the black bars as well?
it just means rendertimes will be low if blackbars are removed and nothing else changes because madVR processes less pixel now. with zooming it may need more processing.

[70MM]

it just means rendertimes will be low if blackbars are removed and nothing else changes because madVR processes less pixel now. with zooming it may need more processing. So one gains in one area but then looses in another. lower render times but higher processing due to the zoom thats required.

[huhn]

you either loose performance and you win it. it depends on the zooming.

if you are presenting a 21/9 movie from a BD on a 16/9 screen and you don't zoom black bars but crop them than you clearly win performance.

[AngelGraves13]

teaser:

Two new algorithms coming in the next build, probably out tomorrow:

blocking (http://madshi.net/deblockA.png) -> cleaned (http://madshi.net/deblockB.png)
mosquito (http://madshi.net/mosquitoA.png) -> cleaned (http://madshi.net/mosquitoB.png)

Yay! Finally...I've been saying we needed a deblocking filter for years now.

What kind of performance hit are we looking at with these enabled?

I'm guessing each setting will have its own checkbox?

[70MM]

I never tick any of the zoom black bars away. I use an Isco anamorphic III lens with motorised masking to 16:9 and 4:3.

[dvd1]

https://s1.postimg.org/6y9b6o3jsb/Immagine.jpg (https://postimg.org/image/6y9b6o3jsb/)
576 to 1080 p
all bilinear
What devrei improve?

[huhn]

if you give me the source 720x576 unscaled image i can show you.

[AngelGraves13]

https://s1.postimg.org/6y9b6o3jsb/Immagine.jpg (https://postimg.org/image/6y9b6o3jsb/)
576 to 1080 p
all bilinear
What devrei improve?

Go fully Jinc if you can. Bilinear is a terrible scaler.

[ryrynz]

Will there be different settings for the deblocker? I'd like something that could be kept enabled for almost all content without losing detail.

[mzso]

I don't understand what he's saying, to be honest.

Well, maybe you can determine if the bug is on progDVB's side or not then.

[nevcairiel]

Will there be different settings for the deblocker? I'd like something that could be kept enabled for almost all content without losing detail.

That would be nice. Enabling things on a per-content basis is annoying and I'm usually too lazy for that.

[x7007]

I did. It let me skip the nvidia 3D Vision Wizard. But if I do the nvidia image test it still crashes the same as the wizard crashed and madVR does not initiate 3D mode. I get no 3D logo on the TV as before. It just switches to 24hz and says that the content is "running in 3D".

I think this has started after latest Windows 10 update.
Reinstalled latest nvidia driver, cleaned with DDU, tried going back to DXVA copy-back... No go. :(

Can anyone try with Windows 10 v10.0.15063 & 385.69 drivers?
The 3D vision wizard still crashes btw, but that's nvidia being nvidia I guess.
To Full Quote https://forum.doom9.org/showpost.php?p=1821102&postcount=46316

3D vision wizard crashes, it's bug in the drivers since 384.xx just wait for a fix.

[nevcairiel]

I did. It let me skip the nvidia 3D Vision Wizard.

I've never once even had to use the wizard to use 3D video in madVR. It should only be required for 3D games.

[FDisk80]

I've never once even had to use the wizard to use 3D video in madVR. It should only be required for 3D games.

Yes, that auto enable/disable 3D setting in madVR works like magic.

[cork_OS]

teaser:

Two new algorithms coming in the next build, probably out tomorrow:

blocking (http://madshi.net/deblockA.png) -> cleaned (http://madshi.net/deblockB.png)
mosquito (http://madshi.net/mosquitoA.png) -> cleaned (http://madshi.net/mosquitoB.png)
:eek: That's... just awesome! Long awaiting features became a reality.

[VAMET]

Dear madshi

Do you plan to add support for Dolby Vision in your future releases?

I would be glad to see such update. Thank you for your great support.

Sincerely

[foozoor]

Two new algorithms coming in the next build

yeah madshi, finally :D:D:D:D

What about the one-pass, faster and better AdaptiveSharpen?
What about an "anime optimized NGU" to compete with RAVU?

:devil:

[aufkrawall]

There's extreme ringing even with anti-ringing enabled? Do you have a screenshot, or ideally a small sample?

That's with default setting of anti-ringing and anti-bloating filters.
https://abload.de/thumb/off4zsun.png (http://abload.de/image.php?img=off4zsun.png) https://abload.de/thumb/onv0sj2.png (http://abload.de/image.php?img=onv0sj2.png)
It's even more distinct in motion. It's the Samsung "Chasing the Light" showcase video.

[ryrynz]

What about the one-pass, faster and better AdaptiveSharpen?


I find AdaptiveSharpen plenty fast already. As far as improvements to it go if it even needs them that likely won't come till after 1.0


What about an "anime optimized NGU" to compete with RAVU?


Doubt any changes to NGU are coming before 1.0 either, if you think RAVU does well in comparison to NGU AA, post some examples.

[dvd1]

if you give me the source 720x576 unscaled image i can show you.
how do i take the original 720x576?
thank you

[sauma144]

Two new algorithms coming in the next build, probably out tomorrow...

Only two! But that's not enough! Why not 5, 10 or 20! :D

I find AdaptiveSharpen plenty fast already. As far as improvements to it go if it even needs them that likely won't come till after 1.0

If it's better, why not?


Doubt any changes to NGU are coming before 1.0 either, if you think RAVU does well in comparison to NGU AA, post some examples.

Will the 1.0 version be available before 2020? :p

[ABDO]

What about the one-pass, faster and better AdaptiveSharpen?
What about an "anime optimized NGU" to compete with RAVU?

:devil:
for me new Long awaiting features appear, better than Improve existing one, especially I'm watching anime with alots of noise and blocking, and such algorithms make watching much better;):D

[Siso]

for me new Long awaiting features appear, better than Improve existing one, especially I'm watching anime with alots of noise and blocking, and such algorithms make watching much better;):D

Yeah, but I guess they will be useful for anime only, not for movies.

[Q-the-STORM]

Generally, in this situation it *seems* that the Nvidia custom mode control panel works, but it doesn't, really. Try rebooting your OS and the custom mode you created through the Nvidia control panel should be gone (it is on my PC). The Nvidia control panel doesn't report failure of the Nvidia custom mode APIs, while madVR does. That's why it seems to work in the Nvidia control panel (but doesn't survive a reboot, or a GPU reset), while madVR properly detects the failure and reports it.


ok, but the nvidia control panel custom modes do work, at least somewhat...

just did some testing, created a custom mode in nvidia, 2160p24, only changed 24.000 Hz to 23.976Hz, left everything else. (So I basically only changed the pixel clock from 297.00mhz to 296.70mhz if I understand nvidias panel)

So I tested standard modes 23 and 24 and my custom 24 mode by just running a 23.976fps video for about 10 minutes each time

in FSE, every mode stabilized at the same Hz:
23.97894 Hz
1 frame repeat every 4.41 minutes

in windowed or fullscreen windowed, I had different Hz every time:

standard 23:
23.97894 Hz
1 frame repeat every 4.41 minutes

standard 24:
23.99914 Hz
1 frame repeat every 41.71 seconds

Custom 24:
23.97489 Hz
1 frame drop every 1.07 hours

So the new custom mode does "something" different than the standard modes.

Interesting thing to note is that according to the madVR custom modes tab, my custom 24 mode is not active, the 23Hz one is. This might be, because my custom 24 mode seems to have the exact same timing parameters as the standard 23 mode, even though I get different results in playback.

I don't know if this is new information, or if this was obvious to you from the start, but there it is...

Since I'm using windows 10, I should be getting 10bit in windowed fullscreen, correct?
If so, 1 frame drop every 1.07 hours is good enough for me and I'm done for now.
Though if bugs get fixed in the future, I'm happy to do more testing.


On an unrelated note:
the "display bitdepth - toggle" shortcut only goes through 9bit and 10bit the first go round... So if I have a 10bit display and use the shortcut, it'll go from 10 bit to 9, 8, 7, 6, then back to 8, 7, 6, 8, 7, 6 etc. Shouldn't it go back up to 10 bit?

[huhn]

how do i take the original 720x576?
thank you

mpc-hc -> dile -> save image -> save as png.

[ABDO]

Yeah, but I guess they will be useful for anime only, not for movies.

it will be useful for both (crappy and dirty) anime and movies :)

[dvd1]

mpc-hc -> dile -> save image -> save as png.
https://s1.postimg.org/7tdc6052ij/dvd_snapshot_00.04.44_2017.10.08_17.41.43.png (https://postimg.org/image/7tdc6052ij/)

tell me
thank you

[huhn]

or just look at this:
source: https://abload.de/img/title06.mkv_snapshot_j1ziv.png

billinear: https://abload.de/img/bilinearwusqx.png
spline: https://abload.de/img/splinewyszw.png
ngu soft: https://abload.de/img/ngusoftveryhighhgsnw.png
ngu standard: https://abload.de/img/ngustandardveryhighgeszx.png
ngu sharp: https://abload.de/img/ngusharpveryhighxnssg.png

the zipper and the hair is very different with the scalers and billinear is just bad. the source is pretty soft so the scaler are not doing much on the background because it is not in focus.

[omarank]

teaser:

Two new algorithms coming in the next build, probably out tomorrow:

blocking (http://madshi.net/deblockA.png) -> cleaned (http://madshi.net/deblockB.png)
mosquito (http://madshi.net/mosquitoA.png) -> cleaned (http://madshi.net/mosquitoB.png)
Wow!! Just superb.

[huhn]

https://s1.postimg.org/7tdc6052ij/dvd_snapshot_00.04.44_2017.10.08_17.41.43.png

tell me
thank you

bilinear: https://abload.de/img/billineare3pa6.png
ngu sharp: https://abload.de/img/ngusharpe1rem.png

i don't even know where to start it is anywhere.

if you can't see the major difference I suggest in a not offensive way to meet a optician to fix the problem.

the difference is really really big al over the image except the sky.

[ashlar42]

Thanks madshi for deblocking and mosquito removal algorithms. Looking forward to them. Always appreciate the work you put in your awesome renderer.

[iSeries]

well working perfectly fine here: https://abload.de/img/workingfine3wukf.png

Thanks huhn, although that isn't happening for me. Zoom functions not working for me with d3d11 copyback. Nvidia 950.

[madshi]

madshi so does this mean if we tick the box and zoom the black bars away, the image would be better again since there is no processing going on of the black bars as well?
That won't affect image quality.

Do you plan to add support for Dolby Vision in your future releases?
I don't know right now. Dolby Vision specs are complicated, it will also need to be supported by the decoder. It's a lot of effort.

What about the one-pass, faster and better AdaptiveSharpen?
Is in v0.92.5.

What about an "anime optimized NGU" to compete with RAVU?
Maybe I'll look into that at some point, but it doesn't have a very high priority right now. With the anime image bjin has selected for his own image comparison NGU AA looks better to me than RAVU. So I'm not sure how much I can optimize there.

That's with default setting of anti-ringing and anti-bloating filters.
It's even more distinct in motion. It's the Samsung "Chasing the Light" showcase video.
Thanks, I'll have a look at that when I find some time.

ok, but the nvidia control panel custom modes do work, at least somewhat...
Have you tried rebooting your OS? Is the custom mode still available after the reboot?

Well, maybe you can determine if the bug is on progDVB's side or not then.
I don't even know what the bug is all about. I don't have the resources to go to other forums to look for bugs. If the ProgDVB developer thinks there's a bug in madVR then he'll welcome to contact me about it. He did contact me about the deinterlacing bug, and it's fixed now.

Will there be different settings for the deblocker? I'd like something that could be kept enabled for almost all content without losing detail.
There is a strength setting, and unfortunately higher strengths can be quite destructive. Well, they most of all destroy block and mosquito artifacts, but eventually also some image detail. I'm not sure, maybe you can leave the lowest strength enabled at all times, you'll have to test if it costs too much detail or not. But I rather recommend having this algorithm disabled by default and only enable it if your eyes complain about compression artifacts during playback.

[madshi]

madVR v0.92.5 released

http://madshi.net/madVR.zip

* added new algorithm "reduce compression artifacts"
* added new algorithm "reduce random noise"
* added file name tag and keyboard shortcut support for new algorithms
* switched to igv's AdaptiveSharpen variant
* custom mode optimization is now available for modes not created with madVR
* fixed: mode optimization didn't work for Nvidia (introduced in v0.92.4)
* fixed: ProgDVB: DXVA deinterlacing didn't activate after channel switch
* fixed: potential freeze when freeing madVR instance
* fixed: playback trouble after switching video files (introduced in v0.92)
* fixed: screenshot memory leak
Notes:

1) Please test the new algorithms and let me know what you think!

2) The block and mosquito noise reduction is actually *one* algorithm named "reduce compression artifacts", not two separate algorithms. The other new algorithm is "reduce random noise".

3) Currently the new algos are only available in one quality/speed variant. I might offer different quality/speed settings in the future. I'm not talking about the algorithm "strength", which is a totally different thing. I'm talking about the speed of the algorithm, and the resulting quality of the noise/artifact reduction.

4) Currently chroma noise/artifact reduction is very slow because I'm doing it in luma resolution. In a future version I'll do it on the original chroma data instead which should reduce the chroma processing hit to a quarter of the current cost for 4:2:0 content.

5) Random noise reduction is currently limited to max 5 strength. In a future version it will probably go up to 10.

6) The higher "reduce compression artifacts" strength settings will totally destroy your images, so DON'T use them. You've been warned! They're there only for videos which are borked beyond good and evil. My advice is to keep this algo disabled by default, or at lowest settings and only increase the strength if really necessary. You can use file name tags (e.g. "some video deblock=5.mkv") to force a specific deblock setting for a specific file.

7) Please compare the new AdaptiveSharpen version (it's igv's variant) to the original AdaptiveSharpen algorithm. On a quick check I liked igv's variant better, but I'd like your feedback on that.

[huhn]

Maybe I'll look into that at some point, but it doesn't have a very high priority right now. With the anime image bjin has selected for his own image comparison NGU AA looks better to me than RAVU. So I'm not sure how much I can optimize there.

what about speed? and is RAVU still work in progress?
if it is faster and still in progress why not add it as long as it is a meaning full alternative.

[ABDO]

madshi you are wonderful, great work, mucht thanks :D:thanks:

[mzso]

I don't even know what the bug is all about. I don't have the resources to go to other forums to look for bugs. If the ProgDVB developer thinks there's a bug in madVR then he'll welcome to contact me about it. He did contact me about the deinterlacing bug, and it's fixed now.
That would require him to acknowledge it as a bug, which so far I couldn't get him to do. (The deinterlacing issue took a year...)
I got nowhere with that. Keeps pointing me towards madVR.

But basically the bug is that I get this:
(madvr)
https://abload.de/img/startwithmadvre6jlu.jpg

Instead of this:
(EVR)
https://abload.de/img/startwithevr8fky4.jpg

(Even though I have HD skin enabled.)

[bitterman]

The decoder queue goes empty in the moment when this happens. Which means that the decoder is not sending madVR any frames in this situation. I'm not sure why, but it doesn't seem like it's likely to be madVR's fault. Have you tried another video renderer like VMR9 or EVR? If the problem also occurs with them, then that would prove that it's not madVR's fault.


It wasn't a fun thing to do, but I've watched the movie with both EVR (custom presenter) and VMR9 (renderless) for approximately 30 minutes straight. Nothing happened.

The moment I changed back to MadVR. The freeze image issue was there again.

The problem only goes away when I untick HEVC in LAV video configuration. Then the playback is smooth and nothing happens.

It's safe to say that the problem is causing because of my not HEVC compatible video card, the GTX 650 Ti.

I have no idea why with EVR and VMR9 the freeze image issue didn't occurred.
Because while I was testing I did had the HEVC box ticked in LAV video and the current active decoder was my gpu instead of avcodec.

No, R7 doesn't support HEVC at all. Then LAV simply falls back to software decoding. Good software decoding is often better than "bad" (no ASIC) hardware decoding.

If you open LAV Video settings during playback it will tell you which decoder is currently active. ("avcodec" = (ffmpeg) software decoder)

Yes, you were right. On my dad's computer the active decoder was "avcodec".
Even though HEVC was still ticked in his LAV video configuration it did automatically went for avcodec instead of using his gpu.

[Ver Greeneyes]

4) Currently chroma noise/artifact reduction is very slow because I'm doing it in luma resolution. In a future version I'll do it on the original chroma data instead which should reduce the chroma processing hit to a quarter of the current cost for 4:2:0 content.
Sounds good. The (old Youtube) videos I'm testing this on right now need a huge amount of deblocking in the chroma channels, but luma only needs some minor adjustments. Would it be possible to have a separate strength setting for the chroma channel processing? On these videos I need a strength setting of 9 or 10 to fix the chroma, but that obviously does far too much damage to the complete image.

[Anime Viewer]

madVR v0.92.5 released

http://madshi.net/madVR.zip

1) Please test the new algorithms and let me know what you think!


Getting a failed error message (repeatedly) trying to download v0.92.5. Anyone else experiencing this problem?

[dvd1]

bilinear: https://abload.de/img/billineare3pa6.png
ngu sharp: https://abload.de/img/ngusharpe1rem.png

i don't even know where to start it is anywhere.

if you can't see the major difference I suggest in a not offensive way to meet a optician to fix the problem.

the difference is really really big al over the image except the sky.
how do you change stairs on an image. ping
thank you

[juanchu]

MPC-BE (1.5.1 (build 2755) beta) crashes with madVR turned on. If I try to change file after playing it or if I try to close player, then crashes. With previous version, this didn't happen.

However, there is no problem with MPC-HC (1.7.13).

[ABDO]

Getting a failed error message (repeatedly) trying to download v0.92.5. Anyone else experiencing this problem?

no,every thing is normal here

[mzso]

madVR v0.92.5 released


* fixed: ProgDVB: DXVA deinterlacing didn't activate after channel switch
* fixed: potential freeze when freeing madVR instance
* fixed: playback trouble after switching video files (introduced in v0.92)
[/code]


I think something may have gotten awry. Now I get a crash every time I change-channel/try-to-change-renderer/unload-tuner with ProgDVB and every time I stop-playback/open-a-new-file-during-playback with potplayer. So it seems like when madVR is unloaded.

I couldn't test for the previous deinterlacing or hanging issues because of this.

(all I did was overwrite the files in the madVR folder and ran instal.bat)

[madshi]

That would require him to acknowledge it as a bug, which so far I couldn't get him to do. (The deinterlacing issue took a year...)
I got nowhere with that. Keeps pointing me towards madVR.

But basically the bug is that I get this:
(madvr)

Instead of this:
(EVR)

(Even though I have HD skin enabled.)
Is there an easy way for me to reproduce this without requiring any actual DVB hardware?

It wasn't a fun thing to do, but I've watched the movie with both EVR (custom presenter) and VMR9 (renderless) for approximately 30 minutes straight. Nothing happened.

The moment I changed back to MadVR. The freeze image issue was there again.

The problem only goes away when I untick HEVC in LAV video configuration. Then the playback is smooth and nothing happens.

It's safe to say that the problem is causing because of my not HEVC compatible video card, the GTX 650 Ti.

I have no idea why with EVR and VMR9 the freeze image issue didn't occurred.
Because while I was testing I did had the HEVC box ticked in LAV video and the current active decoder was my gpu instead of avcodec.
Does madVR show "(DXVA11)" at the end of the 4th line in the Ctrl+J OSD when these freezes occur?

Sounds good. The (old Youtube) videos I'm testing this on right now need a huge amount of deblocking in the chroma channels, but luma only needs some minor adjustments. Would it be possible to have a separate strength setting for the chroma channel processing? On these videos I need a strength setting of 9 or 10 to fix the chroma, but that obviously does far too much damage to the complete image.
Can you make a sample available, or 2 or 3 which show this problem? Would be interesting to look at them. But yes, different strengths for luma and chroma are possible, if really needed.

Getting a failed error message (repeatedly) trying to download v0.92.5. Anyone else experiencing this problem?
Seems to work fine for me.

-------

Just noticed a crash problem in v0.92.5, which happens when switching from one video file to another, or sometimes even just closing a video file. Already found the cause of the problem, stupid bug. Sorry, my bad, will fix that tomorrow in v0.92.6.