madVR - high quality video renderer (GPU assisted)

[imhh11]

There's somebody hooking D3D9 on your PC. I'm not sure who or why. Do you have any funny NVidia extensions installed, like virtual desktop stuff, or maybe some remote desktop software running or something?

like teamviewer ?
I'll try uninstalling geforce

[Ver Greeneyes]

If GeForce Experience is installed, I'd try uninstalling it or turning off its capture capabilities.Yeah, the most recent GeForce Experience turns its "Share" overlay on by default IIRC. You can turn it off in its settings.

[makakam]

Are soften edges and grain enabled by default or do I have to turn them on somehow?

[huhn]

you have to turn them on under upscale refinement.

[Anima123]

Here's the log for debug build: https://www.dropbox.com/s/t541lccfvqnxvgf/madVR%20-%20log.rar?dl=0

And, I noticed that the debug mode does not activate NGU at all, and the OSD works.

[imhh11]

Here's the log for debug build: https://www.dropbox.com/s/t541lccfvqnxvgf/madVR%20-%20log.rar?dl=0

And, I noticed that the debug mode does not activate NGU at all, and the OSD works.

same thing here.

[makakam]

you have to turn them on under upscale refinement.

Thanks, found it at the same time you wrote ;)

[StinDaWg]

madshi, for NGU it is suggested to use Catmull-Rom with anti-ringing and linear light disabled, instead of SSIM for downscaling. Correct?

[aufkrawall]

I personally wouldn't expect this. Less sharper bicubic is more aliased and the ringing by sharper is covered well by AR.

[thighhighs]

What about high vs med quality? For me, med looks better (quality\performance), but i not sure. With gtx690 i cant use high in any cases, but may be i should downgrade other settings, if high vs med add extra quality.

[aufkrawall]

With medium, contours look more aliased and shaky in cartoons than with high.

[Asmodian]

madshi, for NGU it is suggested to use Catmull-Rom with anti-ringing and linear light disabled, instead of SSIM for downscaling. Correct?

No, NGU is tuned to upscale something that was downscaled that way but this doesn't mean you should then downscale its output the same way.

We can upscale with NGU to closely match the original image but then downscale with the best downscaling available, unlike what studios usually do.

[BetA13]

meehh. i just tested it out.. seems very nice i have to say, but perf. wise a nogo for me..

sXBR - 45% gpu usage.
NGU - Medium - 95% GPU usage

really heavy on my gtx 670..

guess i have to update my hardware sometimes... greeetz

[aduwind]

Madshi thank yoyu,
My settings

Epson EH TW5500 Projector
OS W10 64
GPU GTX 1080
Nvidia Power management adaptive
SSD internal MV2
Full blu-ray 1:1 movies1080 p23,976 or p 24
My settings - average render times: 39
PROCESSING
Artifact removal
reduce banding Ticked
default debandig High
Strenght during fade High
reduce ringing arifacts Ticked
Reduce Halos Ticked
Image ehnancement
sharpen edge Uncecked
Crispen edges Uncecked
Thin edge Uncecked
enhance detail Uncecked
LumaSharpen 1,39
ADAPTIVE sharpen Unchecked
Anti bloating Unchecked
Antiringing filter Unchecked
SCALING ALGO
Chroma upscaling NGU Medium
SUPERES Unchecked
Image downscaling
SSIM 100%
Antiringing filter Unchecked
Scale linear light Ticked
bloating filter Unchecked
Image Doubling (X2 always)
Luma NGU High
Chroma Unchecked
Image Upscaling
JNC 100%
Activate antiringing Unchecked
Scale in sigmoidal light Ticked
Upscaling refiniment
Sharpen edges Unchecked
crispen Unchecked
Thin edges Unchecked
enhance detail 1
lumasharpen 1,02
Adaptive Sharpen 1,4
antiringing Filter Unchecked
Superes Unchecked
Superes - linear light Unchecked
Superes -Antiringing Unchecked
Refine image only after~2x resampling Unchecked
Refine image only after upscaling complete Ticked
RENDERING
Smooth motion Unchecked
dithering Ordered
colored noise Ticked
change dither for every frame Ticked

Average rendering for: Only Luma doubling
Crhoma upsc = NGU medium
Imag Upscl = Jnc

ALGO.................. Ms
NNEDI3 256 - 62
NNEDI3 128 - 41
NGU High - 39
NGU MEDIUM 25
NNEDI3 64 30
NNEDI3 32 25

Average rendering for Luma and Chroma doubling:

Algo................................ MS
NNEDI3 256 x NNEDI3 256 /....... 153
NGU High x NGU High- /....... 83
NGU Med x NGU Med /....... 35
NNEDI3 128 X NNEDI 32 /....... 57
NNEDI3 64 XNNEDI3 64 /....... 57
NNEDI3 64 XNNEDI3 16 /....... 40

NNEDI3 32x NNEDI3 32 /....... 39
NNEDI3 32 x NNEDI3 16 /....... 34

[leeperry]

Ok, two RX480 users now with very bad performance results. I smell a pattern. So my current impression is:

1) NVidia: a bit slower than expected.
2) RX480: a *lot* slower than expected.
3) All older AMD generations: just as fast as expected.

Yeah it flies with a HD7850 and the 13.12 drivers on W7, looks quite amazing with soften edges@1 :cool:

OSD no workee though :o

I also like how the new config panel doesn't trigger UAC to ask whether I want to authorize it to run as it always took a few secs to show up.

I was hoping you could chime in regarding windows CLUT's bit-depth when outputting 10bit from mVR but I guess you don't hack things around and let both windows and graphic drivers know that it's 10bit output so there shouldn't be a reason why the CLUT would be 8bit..

[AndreaMG]

For people reporting that install.bat doesn't work in Windows 10:

madshi knows. It's been reported multiple times already.

To work around this you just need to copy the files from the new archive into the folder of the older version, overwriting the old files. I *always* update madvr this way. Since this update method doesn't change where the madvr files are located, it is not necessary to run install.bat.

Ok Thanks :) In my case, using SVP, even medium is too much and my GTX970 OCed cannot handle it, I didn't have much time to test the new algo but sharp wise it seems on par with jinc supersampling tap 3/4 but much more taxing, let's wait for performance improvements otherwise I'll stick with jinc supersampling

[Smithy]

Ordered Dithering, Fullscreen Window Mode (8 backbuffer) D3D9, CPU64/GPU24
no General Settings Checked, no Image Enhancements Checked and no Trade Quality for Performence Checked
reduce Banding High/High, reduce Ringing and dark Halos
Chroma ReconSoft, Image Upscaling Jinc AR SL, Soften Edges 1, Add Grain 1

720x388 25p -> 1080p60 Smoothmotion

http://screenshotcomparison.com/comparison/191181

The add Grain (static) option looks very nice @ PQ with crappy Sources (blocking, artefakts, dnr'd or destroyed grain by low Bitrates).
But changing Grain with every frame is much better in Motion Scene (Dirty Screen effect) !
Any Chance for dynamic option, too ?

without Doubling (7,5 - 8ms)
SXBR AB25 Luma, SXBR AB25 Chroma (9 - 9,5ms)
NGU Luma Med (13,5 - 14ms)
NGU Luma Med, Chroma Mid (15 - 15,5ms)
NGU Luma High (19,5 - 20,5ms)
NGU Luma High, Chroma Mid (20 - 21ms)
NGU Luma High, Chroma High (36ms jumping to 41ms) -> Frame Drops

EDIT:
I have Notice Coil Whine with GPU and @ Display 1 (Monitor) by switsching to Display 2 (TV) with NGU Doubling.
With Monitor Power OFF, the Monitor Sound Frequency are Gone ! ^^

[aufkrawall]

Ok Thanks :) In my case, using SVP, even medium is too much and my GTX970 OCed cannot handle it, I didn't have much time to test the new algo but sharp wise it seems on par with jinc supersampling tap 3/4 but much more taxing, let's wait for performance improvements otherwise I'll stick with jinc supersampling
Could you please elaborate what you do mean with "jinc supersampling"? Something comparable to EWA scalers in MPDN? :)

[Mistery 73]

I apologize I know you have already answered that question 1,000 times on the win 10 problem .. but I can not come out of the problem does not fire me NGU nowhere I also ran the madashi guide posted here on the forum .. thanks

[AndreaMG]

Could you please elaborate what you do mean with "jinc supersampling"? Something comparable to EWA scalers in MPDN? :)
I meant jinc upscaling with SuperRes 3-4. SILLY ME I was testing on my 1440 monitor :eek: now I'm watching BUFFY (576p) on my 1080 plasma TV with the new algo (luma only) set to medium with even some sharpening on top and it looks FABULOUS, no dropped frames in 60fps SVP :) :thanks: Madshi, goodbye NNEDI it was good till it lasted :D

[CruNcher]

Yes, it does. But it's also about a 100x slower than NGU.



I haven't heard from tritical in years. NGU is not based on NNEDI3 in any way. E.g. NNEDI3 needs OpenCL (or D3D11 shaders) to run with decent performance. NGU runs in simple PS3.0 pixel shaders.

Nobody has had a hand in NGU other than me. Well, of course I must have read several dozen scientific papers about image processing and upscaling, and I've certainly taken many good ideas from several such papers, and added a couple of my own ideas.


Doesn't look all that great to me, to be honest.




Proof that your PS 3.0 implementation performs equal or better sorry madshi but that's pretty unbelievable that you even come close to it in all test cases.

And the most insane thing i ever read from you.

I would even go as far and say you don't beat Samsung SAIT R&D in anyway (Super Resolution that's used in every SUHDTV) with that PS 3 shader.

[aufkrawall]

I meant jinc upscaling with SuperRes 3-4. SILLY ME I was testing on my 1440 monitor :eek: now I'm watching BUFFY (576p) on my 1080 plasma TV with the new algo (luma only) set to medium with even some sharpening on top and it looks FABULOUS, no dropped frames in 60fps SVP :) :thanks: Madshi, goodbye NNEDI it was good till it lasted :D
Thx. I'm on WQHD too, and agree that for 1080p -> WQHD Jinc AR + SR3 is better than NGU (faster and sharper).

[Sp00kyFox]

thanks for the new version madshi, the new NGU algorithm is impressive! I noticed a couple of things.

the new add grain feature in "upscaling refinement" looks very stationary. in scenes with a moving camera it looks like as if there were a dusty glass panel on top of the screen instead of the grain being a part of the film material.

I also noticed that the feature "enhance details" seems to be executed after adding the grain which enhances it. the grain should probably come last.

[AndreaMG]

I also noticed that the "add grain" feature is too much stationary at the point that is distracting, the idea behind it is good though (even if I'm not a super fan of grain:) but the implementation must be revised IMHO). @Madshi is your algo similar to this one http://avisynth.nl/index.php/AddGrainC?

[AndreaMG]

Proof that your PS 3.0 implementation performs equal or better sorry madshi but that's pretty unbelievable that you even come close to it in all test cases.

And the most insane thing i ever read from you.

I would even go as far and say you don't beat Samsung SAIT R&D in anyway (Super Resolution that's used in every SUHDTV) with that PS 3 shader.

Yes of course and you might also add that samsung super powerful customchips perform better than a 500 dollars card at full power with madshi algo :D

[Stacey Spears]

I would even go as far and say you don't beat Samsung SAIT R&D in anyway (Super Resolution that's used in every SUHDTV) with that PS 3 shader. Can you provide some examples of content that shows where you like Samsung over NGU? Or frankly, any content to test up-scaling.

To be clear, I am not questioning your opinion. NGU is new and it will improve over time. This thread is already providing great examples for madVR to grow.

I am personally looking for examples of content that I can use in new scaling test pattern(s) I am building for my UHD HDR BD disc. If there is specific content you use I would love to see it. Or any ideas you have. If you are will to share, of course.

In this scenario I am looking for scaling images up. I will provide 720x480, 720x576, 1280x720, 1920x1080 and 3840x2160. Primary focus is scaling everything up to 3840x2160 for playback. The inclusion of 3840x2160 is to have a reference to compare the up-scaled images with.

For post production work, I am also looking for ideas on testing down scaling algorithms. I really like the SSIM scaling in madVR. In post, we are shooting at 5K, 6K and 8K and need to scale down to Full 4K, UHD and Full HD resolutions.

[Sunset1982]

I have a XFX RX 480 GTR Back Edition and I get worse performance.

Any idea why this is? I mean this is one of the fastest RX 480 cards and I would expect similar or slightly better results. Sunset1982 are you using some settings that makes it faster? There is a ratio of approx. 1.5. What can explain this?

I'm at work right now, but I will post my settings tomorrow.

Be careful, use a file equal to mine with exactly 1920x1080p, 23,976, no black borders (mine was a avatar remux with 32mbit) to get comparable results.

[HillieSan]

I'm at work right now, but I will post my settings tomorrow.

Be careful, use a file equal to mine with exactly 1920x1080p, 23,976, no black borders (mine was a avatar remux with 32mbit) to get comparable results.

Thanks. I have used London has fallen, 1920x1080p, 24 fps, 2.39:1

[magic144]

Hi,
I dunno what's happening but I CANNOT run the installer (.bat) for the latest build (0.91).
The regsvr32 command HANGS. (x86 W10 Home system)

Can revert to 0.90.24 no probs.

Any help??

m

[Asmodian]

It is a known issue. Simply replace the old files, there is no need to rerun the installer.

[huhn]

known issue the installer simply doesn't work.
install an old version and copy the new version over it.

[magic144]

ok thanks guys

to clarify:
regsvr32 from an admin cmd prompt seems to work (exits), but is still seen to be running in process list
(apparently, madHcNet32.dll is still in use by it)

if you run the install.bat file (as admin), it hangs

will do as you suggest for this release, use 0.90.24 installer first, then just copy files
never had this issue before :)

*sorry for not reading back far enough in the forum to see others with same issue! Hope madshi figures it out

[strumf666]

Win10x64 i5 8Gb R9 270 crimson 16.11.4 using dx11 NGU-medium/no superres/OD
I am getting around 14ms average render stat for 1080p playback, which makes it borderline usable in conjunction with SVP.
A couple of observations, gpu doesn't go to full clocks during playback (only 900/1400 boost is 925/1400 while I OCed it to 1050/1500 for testing ), while during 3d test or SVP test it does. DX9 doesn't work, just eats all available ram making os response sluggish. Happens in potplayer and MPC-BE.

[Thunderbolt8]

the installer hangs for me with win 8.1 as well so its not solely a windows 10 problem.

[huhn]

Win10x64 i5 8Gb R9 270 crimson 16.11.4 using dx11 NGU-medium/no superres/OD
I am getting around 14ms average render stat for 1080p playback, which makes it borderline usable in conjunction with SVP.
A couple of observations, gpu doesn't go to full clocks during playback (only 900/1400 boost is 925/1400 while I OCed it to 1050/1500 for testing ), while during 3d test or SVP test it does. DX9 doesn't work, just eats all available ram making os response sluggish. Happens in potplayer and MPC-BE.

welcome to the club: http://bugs.madshi.net/view.php?id=431

[strumf666]

Eating the ram, noted, what about the graphic clocks, anybody with AMD graphics noticing the same thing?

[huhn]

well the driver has to select the power state but that isn't working good at the moment: https://abload.de/img/amddriver5wxh3.png

and the option for prefer max performance is gone too.

[strumf666]

It seems r9 270 has an uvd powerstate for playing video; and OCing doesn't affect it...

[huhn]

give software decoding a try.

[markanini]

720p30 to 1080 on my HD7770, enhance details enabled pre-scaling:
SuperXBR: 31ms
SuperXBR AR: 35ms
NEDI: 32ms
NNEDI3-16: 52ms (no chroma)
NNEDI3-32: 75ms (no chroma)
NGU medium: 73ms
NGU high: 146ms

I can avoid dropped frames with SuperXBR-AR by setting usign DXVA2 image downscaling(19ms). I will see if it's a worth it compared to jinc upscaling without doubling for example.

[strumf666]

It seems r9 270 has an uvd powerstate for playing video; and OCing doesn't affect it...

give software decoding a try.
Help with the clocks, I can reach the max clocks now, but overall the playback is worse, less smooth, more dropped frames when ques drop. Ah well, time to decide to either start watching movies on my gaming rig or be content with super-xbr+SR.
Thanks for the idea, anyhow.

[omarank]

Why? 10bit output is an option you can choose yourself. You should know if you have it enabled or not, shouldn't you?
Actually, I wanted to use D3D11 Presentation for 10 bit displays and D3D9 Presentation for 8 bit displays. If you provide “10 bit” variable (or “Output Bit Depth” variable), I can easily create the desired profile rules.


Regarding NGU:

NGU Image Doubling looks very impressive. It’s commendable that you have worked on it entirely.

As others pointed out, there is a dirty look with the “add grain” option. I believe the grains should be dynamic (changing with every frame).

I haven’t had any performance issues or keyboard shortcuts not working problem with the 0.91.0 build on my system with AMD HD8970M card. However, I do notice the loud coil whine with NGU.

Regarding the “feedback”, I think NNEDI3 should stay. I don’t care about SuperRes doubling and NEDI when NGU and NNEDI3 are available.

[feisty2]

madVR v0.91.0 released

5) A word about "ideal upscaling": When trying to upscale a low-res image, it's possible to get the edges very sharp and very near to the "groundtruth" (the original high-res image the low-res image was created from). However, texture detail which is lost during downscaling cannot properly be restored. This can lead to "cartoon" type images when upscaling by large factors with full sharpness, because the edges will be very sharp, but there's no texture detail.


http://www.cs.huji.ac.il/~yoavhacohen/upsampling/

[leeperry]

And it's not my imagination as I double-checked, skipping with NGU in PotP is just as fast as with Jinc....NNEDI3 & sxbr are a lot more sluggish, A/B'ing them against NGU is also laughable, please don't ditch any SR coz I really like them both in combination with NGU and "soften edges"@1. I don't see much point in the new "add grain" static noise option and the older "enhance detail" anymore though.

And so far I still prefer RS soft for chroma with SR@2.

[Asmodian]

http://www.cs.huji.ac.il/~yoavhacohen/upsampling/

That method looks odd:

Our approach extends and refines the patch-based image model of Freeman et al. [10] and interprets the image as a tiling of distinct textures, each of which is matched to an example in a database of relevant textures. The matching is not done at the patch level, but rather collectively, over entire segments. Following this model fitting stage, which requires some user guidance, a higher-resolution image is synthesized using a hybrid approach that incorporates principles from examplebased texture synthesis. We show that for images that comply with our model, our method is able to reintroduce consistent fine-scale detail, resulting in enhanced appearance textured regions.

It is a neat idea but I am not sure attempting to match a low resolution texture to a database of similar high resolution textures and using that texture information to assist upscaling is a good method, their "Our Result" images look very weird.

Maybe a very advanced version could look OK, but unassisted and in real time on video playback is probably a ways off. ;)

[Georgel]

Tested this a little more in depth, and it's true.

It seems that NGU does produce considerably more coil whine.

Most other madVR settings produce no coil whine at all, so it makes me curious why NGU does this..

[Maro3ko]

Hi! Is it somehow possible to manually specify depth of subtitles when playing 3D? I mean external SRT subtitles which does not provide the depth information. Thank you.

[madshi]

OSD no workee though :o
Are you sure that NGU is activated, then? If the OSD doesn't work, probably NGU isn't activated, either (!). You're probably getting super-xbr doubling. It's a known problem, and I'm working on it.

The add Grain (static) option looks very nice @ PQ with crappy Sources (blocking, artefakts, dnr'd or destroyed grain by low Bitrates).
But changing Grain with every frame is much better in Motion Scene (Dirty Screen effect) !
Any Chance for dynamic option, too ?
the new add grain feature in "upscaling refinement" looks very stationary. in scenes with a moving camera it looks like as if there were a dusty glass panel on top of the screen instead of the grain being a part of the film material.
I also noticed that the "add grain" feature is too much stationary at the point that is distracting, the idea behind it is good though (even if I'm not a super fan of grain:) but the implementation must be revised IMHO).
You guys are right, of course. The grain was supposed to move its position from frame to frame. Looks like a bug. I'll fix it for the next official build.

is your algo similar to this one http://avisynth.nl/index.php/AddGrainC?
I don't know.

I also noticed that the feature "enhance details" seems to be executed after adding the grain which enhances it. the grain should probably come last.
You have a good point there. But I wonder if "enhance details" after doubling is really effective? I think if I do "enhance details" first it's better, but I guess any effect will probably just vanish in the grain added for NGU?

I apologize I know you have already answered that question 1,000 times on the win 10 problem .. but I can not come out of the problem does not fire me NGU nowhere I also ran the madashi guide posted here on the forum .. thanks
You mean NGU doesn't get used for you? It's a known problem, I'm working on it.

Can you provide some examples of content that shows where you like Samsung over NGU? Or frankly, any content to test up-scaling.
Hi there, always an honor to see you here!

Proof that your PS 3.0 implementation performs equal or better sorry madshi but that's pretty unbelievable that you even come close to it in all test cases.

And the most insane thing i ever read from you.

I would even go as far and say you don't beat Samsung SAIT R&D in anyway (Super Resolution that's used in every SUHDTV) with that PS 3 shader.
Why are you being so negative/aggressive?

Anyway. Google's RAISR was mentioned in my thread, so I posted my subjective impression. FYI, Shiandow posted a similar impression in the MPDN thread. It's hard to proof anything because as far as I can see, there's no way for me to actually test RAISR. All I can look at is the PDF. They haven't even made the images they used in the PDF available as BMP/PNG files, so we also can't run them through NGU to compare (except those that are well known like the butterfly, see below).

However, I think I have some "weak proof" that I can post:

If you look at page 21 of the Google RAISR PDF, you can clearly see that RAISR has noticeably worse PSNR & SSIM results than SRCNN. And IMHO anything that can't beat SRCNN is kind of underwhelming for a new algorithm. SRCNN was state of the art when it was released but that was a couple of years ago.

Have a look at these two images, with the butterfly also used in the RAISR PDF:

butterfly SRCNN (http://madshi.net/madVR/NGU/butterflySRCNN.png) (PSNR: 32.181)
butterfly NGU (http://madshi.net/madVR/NGU/butterflyNGU.png) (PSNR: 33.689)

Please zoom into these images with your favorite image viewer to see how much better NGU is than SRCNN. And Google RAISR is worse than SRCNN, according to their own PDF!!

Of course it's possible that RAISR could be dramatically faster than NGU, I can't say for sure. But in terms of quality, NGU should be noticeably better than RAISR.

(I should add that I used a custom NGU version which was optimized for best PSNR/SSIM results. I believe all the scientific papers do that. The NGU version used in madVR is tweaked for being as artifact free as possible, which lowers PSNR/SSIM scores but produces much cleaner results. But the NGU version in madVR still beats SRCNN PSNR/SSIM numbers by a noticeable margin.)

Just to be safe: Have you checked the madVR OSD to see if NGU is even being activated in your case? It's a known problem with the current madVR version that on some PCs NGU doesn't activate. In that case super-xbr is used instead. You can verify that by activating NGU for image doubling (not for chroma upscaling!!) and then checking the OSD if it's listed there. Since you don't seem to like what you've seen so far, I can only guess that you're one of those for whom NGU doesn't even activate atm?

Hi! Is it somehow possible to manually specify depth of subtitles when playing 3D? I mean external SRT subtitles which does not provide the depth information. Thank you.
It's not possible atm, I think. I'm not even sure which playback component would be the best to offer such a setting (e.g. subtitle renderer, or subtitle splitter, or madVR, or media player, or movie file splitter etc).

[leeperry]

Are you sure that NGU is activated, then? If the OSD doesn't work, probably NGU isn't activated, either (!). You're probably getting super-xbr doubling. It's a known problem, and I'm working on it.
I've spent a long while comparing them and NGU definitely works as it looks way better for 720p@1080p than Jinc/sxbr/NNEDI3. Cranking it from medium to high also drops frames when using SSIM 2D downscaling, so yes I'd say it works.

Another proof is that "edges soften" and "add noise" also work, PQ is quite amazing on 720p@1080p in combination with RS chroma http://forum-images.hardware.fr/images/perso/ayuluna.gif

[madshi]

Regarding the “feedback”, I think NNEDI3 should stay. I don’t care about SuperRes doubling and NEDI when NGU and NNEDI3 are available.
So you'd like NNEDI3 to stay, but SuperRes and NEDI can go?

http://www.cs.huji.ac.il/~yoavhacohen/upsampling/
Yeah, I've seen 2 or 3 papers like that. But to be honest, while some images look impressive, there's always something ugly in those images, as well. And if these algos fail, they fail spectecularely. I prefer adding random grain over trying to "guess" actual texture detail, because if the guessing goes wrong, we could create some sort of ghost/horror movie instead of upscaling the source video.

Just my 2 cents, of course.