View Full Version : madVR - high quality video renderer (GPU assisted)
digitech
30th October 2016, 05:00
I guess most people don't use the same settings and for a good reason :o. At the same time, my most watched material is drawn and many of it is low resolution or the lineart is a bit thick. Having really strong edge sharpening and thinning helps making the entire image look sharper / like it has a considerably better definition.
Im curious to know which are your settings to enhance anime in sd format, can you share with us?
ryrynz
30th October 2016, 05:19
Im curious to know which are your settings to enhance anime in sd format, can you share with us?
I use a combination of Avisynth and madVR, but regarding madVR, it's all about NNEDI3 really. I use 64 taps for lower resolution stuff with SSIM 1D 100% + Jinc.
madVR's line thinning can be okay when used at low values. I also generally like to use crispen edges at about ~1.0 if I remember rightly.
'Reduce ringing artifacts' is an absolute must for low res content.
Warner306
30th October 2016, 18:40
So it sounds like you don't use this method Warner?
You set your video card to 2160 and only use Image Enhancements for sharpening, is that correct?
If I have this right, why do you prefer the above method not using Image Doubling?
I like to use image sharpening, and image sharpening is not as effective when supersampling (image doubling of a native source). Adding unnecessary image upscaling may actually make the image appear worse.
1.5-2x upscales are where I employ image doubling (SD -> 1080p and 720p -> 1080p).
JarrettH
30th October 2016, 20:27
Hope you guys didn't jinx development by sharing that 4chan thread :D:p
70MM
31st October 2016, 19:33
I like to use image sharpening, and image sharpening is not as effective when supersampling (image doubling of a native source). Adding unnecessary image upscaling may actually make the image appear worse.
1.5-2x upscales are where I employ image doubling (SD -> 1080p and 720p -> 1080p).
So you don't use image doubling on 1080p content, you set the video card to 4K and use image enhancements.
You only use image doubling on DVD and 720p is that correct?
I don't think I have this right actually as if the card is set to 4K everything is 4K....
Asmodian
31st October 2016, 19:37
1.5-2x upscales are where I employ image doubling (SD -> 1080p and 720p -> 1080p).
1.5-2x just means scaling by 1.5 or 2x, e.g. 1280x720 to 1920x1080 is upscaling by 1.5 (720*1.5 = 1080).
CarlosCaco
1st November 2016, 00:42
So you don't use image doubling on 1080p content, you set the video card to 4K and use image enhancements.
You only use image doubling on DVD and 720p is that correct?
I don't think I have this right actually as if the card is set to 4K everything is 4K....
In the specs from JVC dla-x9000 says that the projector it's not true 4K native but have a e4 shift technology that generates/emulates a 4K picture... here's the explanation:
http://eu.jvc.com/microsite/eu/dla-x9000/feature01.html#4k
Also you are using lumagen that accept any input resolution and if you want the processor scale the input to any resolution up to 4K...
So you have a sort of options to choose...
To me seens you have to decide what is best:
Your projector 1080p native resolution or this 4K feature that seens to be high quality (i never saw this feature so i don t know the result)
When you choose what's the best quality
You have to put the gpu and the lumagen to the same resolution so only madvr will scale the image...
Doubling in madvr is good, but depending on resolution it's a matter of taste...
In my opinion doubling a 1080 source to 4K its essencial and will be best that any other upscaling algo
But for a 720 source to a 1080 output there people that like doubling it to 1440 and downscaling to 1080 and people who see no difference to upscaling with jinc using refinements
It's a matter of taste and power...
Also with supersampling... just stay in mind that at 4K your projector its using an algorithm to process the image too
In your situation it's hard to determine what's the right setting... the only thing I can higly recommend is to set gpu and lumagen to the same resolution because doing this you garante that lumagen will not scale the image...
Enviado do meu iPhone usando Tapatalk
Warner306
1st November 2016, 01:03
So you don't use image doubling on 1080p content, you set the video card to 4K and use image enhancements.
You only use image doubling on DVD and 720p is that correct?
I don't think I have this right actually as if the card is set to 4K everything is 4K....
I have a 1080p display.
huhn
1st November 2016, 02:40
In the specs from JVC dla-x9000 says that the projector it's not true 4K native but have a e4 shift technology that generates/emulates a 4K picture... here's the explanation:
http://eu.jvc.com/microsite/eu/dla-x9000/feature01.html#4k
Also you are using lumagen that accept any input resolution and if you want the processor scale the input to any resolution up to 4K...
So you have a sort of options to choose...
To me seens you have to decide what is best:
Your projector 1080p native resolution or this 4K feature that seens to be high quality (i never saw this feature so i don t know the result)
if the lens native resolution is 1080p and it is doing 4 shifts per image the image resolution is true UHD. it is just losing a lot of brightness. kinda like a CRT.
Mr. Durden
1st November 2016, 21:31
Hi everyone,
I'm just curious. Does anybody use the "black bar detection" feature for automatic screen masquerading in your projector setups? MadVR reliable detects the actual aspect ratio and also sends this info optional to the media player (JRiverMC in my case). So it might be possible to adjust motorized screens to respective aspect ratio. I tried jriver http commands, but I don't think this info is available there. In madVR OSD the aspect ratio is shown, but I don't know how to get hands on this info. Maybe someone could give me a hint. It would be great if mediaplayer or better madVR could send UDP Broadcast with current movie aspect ratio. This might be a very cool additional feature for madVR.
Any thoughts on this?
best regards from Berlin,
Stefan
huhn
1st November 2016, 23:09
madVR already support this.
it can control the anamorphic lens of your projector.
go to device -> "your projector" set it to digital projector go to properties set the ip address of your projector.
the rest is under screen config
Mr. Durden
2nd November 2016, 00:26
:-) ah ok thanks, I wasn't really aware of this. I have a Sony projector with support for SDAP, SDCP, DDDP and Crestron Room View protocol. I will have a look into this, maybe a place to start for me.
But rather than controlling the projector lens I wan't to control my motorized screen. I don't want to scretch/compress the original picture. I have a screen with a motorized black border, which can be controlled via IR resp. my home automation system (http get, UDP broadcast, e.g.). It would be much easier to control if madVR would have a option to just UDP Broadcast aspect ratio.
greetings,
Stefan
Dodgexander
2nd November 2016, 21:41
I've tried to search in the thread but couldn't find anything.
Any idea why my present queue is always at 0 or 1/8?
The strange thing is, I am running in fullscreen windowed mode and if something overlays the screen (such as right clicking or having a windows explorer window open in front of the player) the queue fills. Close it so the video is at front again and the queue disappears again.
I am not using fullscreen exclusive mode.
I have reset settings to default in madvr
I have tried setting everything to bilinear in madVR and the same happens
I have updated MadVR to the latest version.
It never happened before on older versions so I'm not sure what has changed.
There are rarely frame drops but its kind of weird
I am using a 576p 25fps test file upscaling to my laptops resolution of 1600x800. I have a i3230M with intel HD4000 graphics but I have also tried using the cuvid decoder in LAV video from the laptops Geforce 650m. The same occurs.
http://i.imgur.com/Q3DXJl2.png
ryrynz
2nd November 2016, 22:28
Any idea why my present queue is always at 0 or 1/8?
Do you use F.lux or something similar? Close all non essential apps to find the culprit.
Dodgexander
2nd November 2016, 22:45
Do you use F.lux or something similar? Close all non essential apps to find the culprit.
Thankyou very much. f.luxx. You're right. I can't believe I didn't think of that especially when it was mentioned already earlier in the thread.[emoji106]
Sent from my Nexus 5 using Tapatalk
Ver Greeneyes
3rd November 2016, 10:15
Why does F.Lux do that anyway? It should just be periodically setting the LUT being used for the display, right? I have a little homemade program to keep my LUT active (some games deactivate it when they open), and it doesn't cause problems with madVR.
ryrynz
3rd November 2016, 11:20
I think it must have something to do with the layered window it uses. Switching to safe mode or switching to movie mode during playback also prevents the present queue from staying at 0-1.
mario_O
3rd November 2016, 15:41
If I enable 'full screen exclusive mode' when I'm watching a film, and then press 'Esc' or 'F11' to go to the desktop or change some settings, I get a black screen. I can hear the audio but I can't do anything else. The only way to close MPC-HC without getting the black screen is by hitting alt+f4. Using latest version and latest Nvidia drivers. Any ideas?
edit- nevermind, I was using an outdated version of madvr. Didn't know the latest version of MPC-HC used an old version of madvr (2014 i think). Dowloaded the latest madvr and everything works great.
tony359
3rd November 2016, 21:07
Hi,
Any help on my query on black bar detection which is not working on my system?
See http://forum.doom9.org/showthread.php?p=1784168#post1784168
Thanks
Tony
huhn
3rd November 2016, 21:22
it doesn't work with DXVA native.
tony359
3rd November 2016, 21:44
Thank you! I'll try using another mode!
pirlouy
4th November 2016, 00:39
Lol he didn't even mention he was using DXVA native. Huhn, you're really on fire ! :D
Dodgexander
4th November 2016, 05:29
Yeah f.lux seems to run in the background and even when it's not doing anything and interfere with the present queue.
On another note has anyone figured out how to avoid the task bar in windows 10 always being on top of windows when playing mpc hc in always on top windowed mode with MadVR? Sometimes I try to watch video in a small window in front of a fullscreen windowed game and it's annoying me that it places always the task bar in front when playing video.
Sent from my Nexus 5 using Tapatalk
XTrojan
4th November 2016, 21:55
Only me that notices ghosting with 3D external subtitles?
Also, can 3D only be run in DX11 mode and exclusive fullscreen?
huhn
4th November 2016, 23:33
you need dx11 for frame packed 3D FSE is not needed if i remember correctly.
the alternative 3D mode should work without DX11 but they are halve resolution.
XTrojan
5th November 2016, 10:45
you need dx11 for frame packed 3D FSE is not needed if i remember correctly.
the alternative 3D mode should work without DX11 but they are halve resolution.
Ah, now I get it.
Is there a reason though why "Ghosting" appears on external subtitles? I'm using xysubfilter, the movie also seems a lot darker than usual.
huhn
5th November 2016, 10:58
sorry i don't know about the ghosting. 3D on most TVs was far from perfect maybe it is crosstalk.
if your screen is an active 3D screen than it is normal that the brightness drops.
ryrynz
5th November 2016, 11:29
Sorry guys, my new algo needs some more tweaking time. Maybe 1-2 weeks more, I don't know.
Is it done yet? The anticipation is killing me.. I need something, anything.
madshi
5th November 2016, 11:54
Fair enough. Unfortunately I need more time, maybe a week, I don't know. I'll give you a little teaser, though, to ease the waiting. The new algo is probably going to be called "NG1" (next generation version 1). It's my own algorithm. It's not finished, but image quality will probably not change much, anymore.
original (http://madshi.net/madVR/AmericanDad/AmericanDadOrg.png) -- | -- Jinc AR (http://madshi.net/madVR/AmericanDad/AmericanDadJinc.png) -- | -- super-xbr 100 (http://madshi.net/madVR/AmericanDad/AmericanDadSuperXbr.png) -- | -- NNEDI3 256 taps + SuperRes 4 (http://madshi.net/madVR/AmericanDad/AmericanDadNNEDI256SR4.png) -- | -- NG1 (wip) (http://madshi.net/madVR/AmericanDad/AmericanDadNG1.png) -- | -- waifu2x (http://madshi.net/madVR/AmericanDad/AmericanDadWaifu2x.png)
FWIW, waifu2x is much too slow for realtime use, I only added it as a reference point for image quality. I've picked this American Dad image because it shows big differences between the various upscaling algorithms. With natural images the difference is smaller.
I'll not answer any questions at this time.
Gagorian
5th November 2016, 12:34
Fair enough. Unfortunately I need more time, maybe a week, I don't know. I'll give you a little teaser, though, to ease the waiting. The new algo is probably going to be called "NG1" (next generation version 1). It's my own algorithm. It's not finished, but image quality will probably not change much, anymore.
original (http://madshi.net/madVR/AmericanDad/AmericanDadOrg.png) -- | -- Jinc AR (http://madshi.net/madVR/AmericanDad/AmericanDadJinc.png) -- | -- super-xbr 100 (http://madshi.net/madVR/AmericanDad/AmericanDadSuperXbr.png) -- | -- NNEDI3 256 taps + SuperRes 4 (http://madshi.net/madVR/AmericanDad/AmericanDadNNEDI256SR4.png) -- | -- NG1 (wip) (http://madshi.net/madVR/AmericanDad/AmericanDadNG1.png) -- | -- waifu2x (http://madshi.net/madVR/AmericanDad/AmericanDadWaifu2x.png)
FWIW, waifu2x is much too slow for realtime use, I only added it as a reference point for image quality. I've picked this American Dad image because it shows big differences between the various upscaling algorithms. With natural images the difference is smaller.
I'll not answer any questions at this time.
Looking very good imo. Can't wait to try it :)
It's clearly better than NNEDI3 with superres, so I guess SR would no longer be needed. Hopefully performance allows upscaling from 720-1080p to 4k on a GTX 1070.
ryrynz
5th November 2016, 12:36
That looks sweet. I'd like to see it vs NNEDI3 256 neurons maybe without SR4. I know you're not answering questions but it looks like reduce ringing isn't enabled for NNEDI3 either that or SR4 is doing that.
Not really a fair comparison with that much haloing going on IMO. Anyway, this looks impressive, thanks!
Even better than Waifu2x (it's just sharper and introduces artifacts) So I'll definitely use this NG1, looks very pleasing.
madshi
5th November 2016, 12:40
The original resolution image is there, so you can create comparison images yourself, with any combination of settings you like.
Sunset1982
5th November 2016, 13:23
holy sh..!!! that looks good! Thank you for that teaser!! Can't wait to get my hands on the next madvr version!
Plutotype
5th November 2016, 14:39
@madshi
Is NG1 somehow comparable to NNEDI3 64 neurons in terms of GPU power required for improvements that you have demonstrated here? Thanks
sauma144
5th November 2016, 15:04
Fair enough. Unfortunately I need more time, maybe a week, I don't know. I'll give you a little teaser, though, to ease the waiting. The new algo is probably going to be called "NG1" (next generation version 1). It's my own algorithm. It's not finished, but image quality will probably not change much, anymore.
original (http://madshi.net/madVR/AmericanDad/AmericanDadOrg.png) -- | -- Jinc AR (http://madshi.net/madVR/AmericanDad/AmericanDadJinc.png) -- | -- super-xbr 100 (http://madshi.net/madVR/AmericanDad/AmericanDadSuperXbr.png) -- | -- NNEDI3 256 taps + SuperRes 4 (http://madshi.net/madVR/AmericanDad/AmericanDadNNEDI256SR4.png) -- | -- NG1 (wip) (http://madshi.net/madVR/AmericanDad/AmericanDadNG1.png) -- | -- waifu2x (http://madshi.net/madVR/AmericanDad/AmericanDadWaifu2x.png)
FWIW, waifu2x is much too slow for realtime use, I only added it as a reference point for image quality. I've picked this American Dad image because it shows big differences between the various upscaling algorithms. With natural images the difference is smaller.
I'll not answer any questions at this time.
Ah thank you very much! Very very very good result! :)
I think you should reconsider the name.
Why not a "madSomething" like name? ;)
Thunderbolt8
5th November 2016, 15:07
the new algorithm looks better than nnedi3. waifux2 still looks best though. too bad, I guess its not possible to make waifux2 faster?
imhh11
5th November 2016, 15:54
wow It definitely looks better than NNEDI3!
I hope my gtx1070 (maybe my gtx960 too) will be able to use the new algo.
thank you madshi, you are a genius.
aufkrawall
5th November 2016, 16:15
Wow, a new scaling algorithm, and it probably outperforms NNEDI3 + SR?
Seems like it's very sharp and artifact free at the same time. Since the contours are not bloated, hopefully the same will be true for naturalistic content.
If yes: You did it. I'll be really glad to get rid of the ringing in most web videos (if the GPU costs aren't much higher than NNEDI3 + SR).
AngelGraves13
5th November 2016, 19:08
The new algo looks great! Very sharp and clean. It'll look even better with SuperRes.
I hope it will be available for Chroma and Image Downscaling as well.
Some other ideas for the name: madVS (mad video scaler)
huhn
5th November 2016, 19:28
i wonder why waifux2 and NG1 are changing the red color...
madshi
5th November 2016, 19:31
Please ignore the color change, that's just a gamut/matrix mismatch, I think.
AngelGraves13
5th November 2016, 19:35
Please ignore the color change, that's just a gamut/matrix mismatch, I think.
Can you please post a photo with SuperRes 4 with NG1? Should really consider a name change for that algo. I recommended one 2 posts above ;)
cyberscott
5th November 2016, 19:44
Fair enough. Unfortunately I need more time, maybe a week, I don't know. I'll give you a little teaser, though, to ease the waiting. The new algo is probably going to be called "NG1" (next generation version 1). It's my own algorithm. It's not finished, but image quality will probably not change much, anymore.
original (http://madshi.net/madVR/AmericanDad/AmericanDadOrg.png) -- | -- Jinc AR (http://madshi.net/madVR/AmericanDad/AmericanDadJinc.png) -- | -- super-xbr 100 (http://madshi.net/madVR/AmericanDad/AmericanDadSuperXbr.png) -- | -- NNEDI3 256 taps + SuperRes 4 (http://madshi.net/madVR/AmericanDad/AmericanDadNNEDI256SR4.png) -- | -- NG1 (wip) (http://madshi.net/madVR/AmericanDad/AmericanDadNG1.png) -- | -- waifu2x (http://madshi.net/madVR/AmericanDad/AmericanDadWaifu2x.png)
FWIW, waifu2x is much too slow for realtime use, I only added it as a reference point for image quality. I've picked this American Dad image because it shows big differences between the various upscaling algorithms. With natural images the difference is smaller.
I'll not answer any questions at this time.
Very interesting. I did notice a small contrast or color shift between the the NNED13 and the NG1 pictures. Look specifically at the red dress or the drapes. In the NG1 processed capture, the color red stays identical to the original while the NNED13 "brightens" the red dress slightly. So, to my eyes, NG1 stays true to the original as far as colors are concerned. Looking forward to trying NG1 when up-scaling to 4K. :)
huhn
5th November 2016, 20:04
so here are my thoughts on the new algorithm. i judged these images on a bad 6 bit nt panel because there is so proper way to judge such a small image on a UHD screen.
NG1 is removing some lines on the wood wallpaper thingy which looks strange. waifux2 does this too. but the source resolution is low.
all x2 algorithm have problems with the top white part of the red dress and NG1(i guess this is a 2x algo too) has this problem too. superXBR is really bad there.
waifux2 is sharper but looks "wrong" the powerswitch looks "deformed" NG1 does this better. i would say that NG1is clearly better in this image than waifux2 at least to me.
waifux2 terrible fails at the necklace just saying...
the name NG1 is fine if you ask me. no "super/mega" stuff no waifu stuff...
Shiandow
5th November 2016, 20:18
Fair enough. Unfortunately I need more time, maybe a week, I don't know. I'll give you a little teaser, though, to ease the waiting. The new algo is probably going to be called "NG1" (next generation version 1). It's my own algorithm. It's not finished, but image quality will probably not change much, anymore.
original (http://madshi.net/madVR/AmericanDad/AmericanDadOrg.png) -- | -- Jinc AR (http://madshi.net/madVR/AmericanDad/AmericanDadJinc.png) -- | -- super-xbr 100 (http://madshi.net/madVR/AmericanDad/AmericanDadSuperXbr.png) -- | -- NNEDI3 256 taps + SuperRes 4 (http://madshi.net/madVR/AmericanDad/AmericanDadNNEDI256SR4.png) -- | -- NG1 (wip) (http://madshi.net/madVR/AmericanDad/AmericanDadNG1.png) -- | -- waifu2x (http://madshi.net/madVR/AmericanDad/AmericanDadWaifu2x.png)
That's an impressive improvement on the previous methods.
Warner306
5th November 2016, 20:21
I like the new algorithm but wonder about the performance cost. Is it accessible to a GTX 750 Ti with a 1080p display?
madshi
5th November 2016, 20:23
That's an impressive improvement on the previous methods.
Thank you... :)
digitech
5th November 2016, 20:31
The new Madshi upscaler algo is the best i have seen in months here and years against other scalers
MariaX9
5th November 2016, 21:17
Looks very good, can't wait for the release. Madshi always delivers. *thumbsup
ryrynz
5th November 2016, 21:24
I like the new algorithm but wonder about the performance cost. Is it accessible to a GTX 750 Ti with a 1080p display?
Probably at the limits at that card but it depends on all your other settings at most importantly what resolution the content is that you're starting off with. I upgraded to a 960 and it was worth while, you may wish to consider that or a 1060. I picked up a 960 for the same price as I sold my 750 for so it was a good upgrade.
NG1 is removing some lines on the wood wallpaper thingy which looks strange. waifux2 does this too. but the source resolution is low.
Not seeing this at all. Please point it out.
Very interesting. I did notice a small contrast or color shift between the the NNED13 and the NG1 pictures. Look specifically at the red dress or the drapes. In the NG1 processed capture, the color red stays identical to the original while the NNED13 "brightens" the red dress slightly.
That won't be NNEDI3, that will be SR4 which is why I asked for the NNEDI3 image alone. I don't have enough time to post it but you should run the original through madVR with NNEDI3 256 and compare to that instead.
vBulletin® v3.8.11, Copyright ©2000-2026, vBulletin Solutions Inc.