madVR - high quality video renderer (GPU assisted)

[ashlar42]

Fair enough. So you'd also like super-xbr to stay (or rather come back)?


Yes, I think we really do need some good compression artifact reducer.Personally I think that the former should depend on the latter. When/if a good compression artifact emerges we could probably do without super-xbr. Till then... I'd keep it in.

[Crimson Wolf]

NVIDIA driver 376.09 is out and it seems to fix the low power mode on optimal and adaptive power settings on my GTX980. YMMV.

[JarrettH]

Not sure why it's a lot faster for you, it was only supposed to be slightly faster. But I won't complain.

To give you some idea, I am running ngu-med with downscale quality medium on a GeForce 550 Ti and Core 2 Duo E6600. Before I had to make a rule to limit ngu to a maximum frame height of 600 pixels. SuperRes covered the >600 pixels because ngu was not fast enough. Now I deleted the SuperRes rule because ngu is fast enough for any 720p to 1080p. The only setup difference I see is in downscaling (catmull prior versus bicubic150AR now, but that's hardly a difference)

Here's a random capture of worst case (25 fps, 1280x720). It's not dropping frames, I was just messing about in and out of full screen.

https://c2.staticflickr.com/6/5337/31155571802_c92819855d_o.png

Most of the time it's in the range of 24 - 30 ms depending on frame height. :cool::thanks:

[Razoola]

Where has the image doubling options gone, I cannot find them anywhere now.

[Crimson Wolf]

NVIDIA driver 376.09 is out and it seems to fix the low power mode on optimal and adaptive power settings on my GTX980. YMMV.

Well, spoke too soon. VP9 1080p60 stuck at lower power mode... both optimal and adaptive. :confused:

Back to max power.

[sneaker_ger]

Where has the image doubling options gone, I cannot find them anywhere now.
Go "image upscaling" and select NNEDI3 or NGU.

http://forum.doom9.org/showpost.php?p=1787605&postcount=40816

[Razoola]

Yea I read that but did not understand that the page was removed totally. Ok I found where the options are now. Is there any reason why the 'chroma quality' drop down box is greyed out when NGU is enabled and is only available when NNEDI3 is used?

[makakam]

With the newest madvr version and my gtx960 I get better rendering times along with NGU very high (chroma upscaling) than I had with NGU high on version 0.91.0 Great!

[Gagorian]

@ madshi

I have tried the newest madVr version and I miss the ability to set the upscaling settings like before. With the max awailable NGU settings, it does'nt fully utilize my GTX 1080 card anymore. This is my settings (with Chroma Upscaling set to NGU-VeryHigh) :

http://bildr.no/thumb/NTFHK2Nx.jpeg (http://bildr.no/view/NTFHK2Nx)

Here is a OSD screen of a 480p, 29 fps clip :

http://bildr.no/thumb/SjljMzJJ.jpeg (http://bildr.no/view/SjljMzJJ)

Max 8.34ms rendering... I could have used much higher settings, but with newest madvr, that is not possible.

Here is a OSD screen of a 720p, 23 fps clip :

http://bildr.no/thumb/bU9vRjJ2.jpeg (http://bildr.no/view/bU9vRjJ2)

Same here. Am I missing something here? My image quality was better with 0.91.1 version of madvr with my (higher) settings there.

What settings were you using that were higher quality? The idea of absolutely maxing out the settings just so the GPU is under higher load doesn't seem very useful. Neither does that necessarily provide any discernible difference in image quality.

[MS-DOS]

madshi, anti-ringing checkbox in image downscaling is greyed out for me. Switching to chroma upscaling and back makes it active until I close the settings window.

[bitterman]

@madshi

In "Chroma upscaling" the default setting is now "Bicubic60" while it was 75 before.

Why did this change? Is it better to leave it at 60 now?

Thanks.

[StinDaWg]

madshi, anti-ringing checkbox in image downscaling is greyed out for me. Switching to chroma upscaling and back makes it active until I close the settings window.
Noticed this too.

[Betroz]

What settings were you using that were higher quality? The idea of absolutely maxing out the settings just so the GPU is under higher load doesn't seem very useful. Neither does that necessarily provide any discernible difference in image quality.

With verson 0.91.1 I was using NGU-High for both Chroma Upscaling and image doubling/quadruple, SSIM2D100AR for downscaling and JincAR for image upscaling. Overkill maybe, but it was sharper and rendertime was low enough.

I prefer to have the option to choose what is used. I do not want to use Bicubic60 AR for any luma upscaling. Well I did notice a differense in IQ with my 480p videos. I reverted back to version 0.90.24 for the time being, and my trusted and loved NNEDI3 settings ;)

[Schwartz]

I've been of same opinion, what you need to do is remove sharpeners superres, some antiblaot settings, everything in image enchantment and image refinement except add grain and maybe thin edges. And it looks without artifact and sharp enough.
NGU conflicts with a lot of things like that creating artifacts. Play with settings more.

I don't use image enhancers -at all-. I only run sXBR 50 for chroma, sXBR 75 for image doubling, then Mitchell for upscaling. Which is a pretty balanced combo that produces softer images with very muted artifacts. IMO, if an algorithm 'overshoots' and then needs to be fixed afterwards through these settings, it's not as good as an algo that doesn't. But take that with a grain of salt. ;)

NGU still oversharpens noticeably on mid-quality SD content. I don't want to resort to enhancements to mess with the image, I'd like to use a clean image doubler instead. But by your argument, I'm already doing it right and NGU should do well. It's simply too harsh of an upscaler.

[aufkrawall]

NVIDIA driver 376.09 is out and it seems to fix the low power mode on optimal and adaptive power settings on my GTX980. YMMV.
Still prevents upclocking here with borderless window + opened Firefox at the same time with "optimal performance" for MPC.

[omarank]

Well, I use CR AR LL to downscale doubled 720p@1080p and SSIM2D100LL+50%AB to downscale SD, this isn't possible anymore?

Why not using the "image downscaling" settings when downscaling doubled upscale?
Not sure I'm crazy about the limited presets for downscaling after image doubling either. I always used AR+LL for downscaling, but you are saying with these presets that LL is "wrong" or that it is "placebo"? I can see the difference between LL on and off, especially with white text, but I don't know which is "correct". I thought the consensus was that you were supposed to use LL for downscaling?
I too have been using Catmull-Rom AR LL for downscaling after 720p to 1080p doubling, with previous madVR builds. Bicubic150 or SSIM1D look too sharp to my eyes in some cases.

Power users end up disappointed with the lack of options (compared to the earlier builds) and normal users never have enough knowledge to not be confused with the remaining options.
I agree.

I would want to choose the upscaling algorithm to be used after image doubling for further upscaling to the target resolution. That algorithm may also be used for upscaling to lower scaling factors which don’t activate the selected image doubling algorithm. For instance, if NNEDI3 or NGU is selected in upscaling settings to be used, say, when scaling factor is 1.5x or bigger, Bicubic60 AR is used for upscaling to less than 1.5x scale factor whereas I would want to use Jinc there.

[Neo-XP]

Reverting back to v0.91.1, because I can't set the luma to > NGU-med < Lanczos3 AR anymore. I was the best quality I could get with my CG. The new luma > NGU-high < SSim 1D100 AR produce a lot more of noise and artefacts.

NGU is epic to upscale FHD content to UHD when adding grain. Best quality with very little artifacts.

Edit : same with Chroma, now we get predefined choices when using NGU. I thought the "automatic" quality selection used the settings from the already defined algorithms in "chroma upscaling", "image downscaling" and "image upscaling" menus. That would make more sense and allow more customization. I don't know what "automatic" stands for there, but if you want to keep it, why not add a "user defined" choice for advanced users ?

[cork_OS]

KNLMeansCL is a good denoiser, I think, but I'm not sure it's the right algo for compression artifact removal.
Well, KNLMeansCL used in Oyster (http://forum.doom9.org/showthread.php?t=173470) (compression artifacts suppressor). Oyster itself is extremely slow, but maybe it can be used as concept or reference like waifu2x.

I wouldn't be surprised if this was perhaps a little higher on madshi's todo list. I have no doubt we'll see something in future.
I agree with you, a good deblocker (either from avisynth or integrated in Madvr), is now the last missing part to achieve a perfect playback chain.
It seems that the madshi heard us.

[JarrettH]

If you choose downscale quality: low, madshi has left you with bilinear, which should hide source artifacts. I think he chose this intentionally to satisfy the SD people.

[aufkrawall]

Ok, I probably have been lead a bit astray by paranoia.
To put it short: I'm fine with the new doubling dialogue, turns out it quadruples the lighttower picture when setting NGU to always activate as supersampling (which is no problem since the setting is only for a profile with high scaling factor).

All the changes of the latest build together allow me to do luma NGU high scaling + chroma NGU low + SSIM for 1080p -> WQHD with my 1070, which is a very good thing.
Edit: Talking about 60fps, ofc.

[pirlouy]

FWIW I'm ok with the new interface config. The "automatic" option is a really nice idea for casual users like me.

[huhn]

If you choose downscale quality: low, madshi has left you with bilinear, which should hide source artifacts. I think he chose this intentionally to satisfy the SD people.

i don't think that helps...

and SD people? i choice the stuff i watch based on the content not on resolution...

[leo91531]

test@GTX1080 (RX480 not yet)
http://i.imgur.com/IA6h3Ld.png

This version feels more simple to adjust:cool:, but AMD's graphics card seems to have some problems.

madvr@RX400 bug?
https://youtu.be/QlVLuGVmZjo
AMD GPU load is unstable,GPU load more than 90%(not Avg load) then video easy to drop frame.(not power saving mode)

RX480 GPU load
http://imgur.com/8Hdz6zz.png

GTX1080 GPU load
http://imgur.com/j7i00JZ.png

[aufkrawall]

madvr@RX400 bug?
https://youtu.be/QlVLuGVmZjo
AMD GPU load is unstable,GPU load more than 90%(not Avg load) then video easy to drop frame.(not power saving mode)
Did you try software decoding? Had a similar issue with a R9 390 and 1080p 60fps H.264.

[Warner306]

I vote to remove downscaling quality or at least add an option to use the downscaling algorithm.

It was actually easier to understand before. Chroma quality is the only box that should be available in a subgroup called "image upscaling." I don't think downscaling belongs. This adds complexity.

Perhaps the image downscaling setting can appear as automatically selected when hovering over NGU or NNEDI3? This avoids a situation where the same setting is repeated.

I am of course assuming image downscaling is part of a profile dedicated to image doubling.

[ikarad]

nearly every g-sync monitor is HFR so just use smoothmotion.
smoothmotion at 120 hz or more is a totally save.
the audio clock mismatch is fixed this way too.

At 120 hz, smoothmotion is useless, not?

[Neo-XP]

I vote to remove downscaling quality or at least add an option to use the downscaling algorithm.

It was actually easier to understand before. Chroma quality is the only box that should be available in a subgroup called "image upscaling." I don't think downscaling belongs. This adds complexity.

Perhaps the image downscaling setting can appear as automatically selected when hovering over NGU or NNEDI3? This avoids a situation where the same setting is repeated across two downscaling menus.

Agreed. I just want to be able to use Lanczos3 AR again for luma/chroma upscaling and luma downscaling, after NGU. The new settings don't give me the same result as before. This is a regression for advanced users.

The low/medium/high/very high quality settings should be chosen as presets in the algorithms menus, not in the NGU part. NGU should be an option in those menus (a checkbox), not used as a replacement, but as a complement.

The "chroma upscaling", "image downscaling" and "image upscaling" menus are totally useless now when using NGU for luma upscaling (because of the forced presets). I don't understand this decision.

[huhn]

At 120 hz, smoothmotion is useless, not?

no not at all.

most content is 23p (24000/1001 hz) and 25/50 content will judder at 120 hz.
smoothmotion is very useful for 120 HZ screen and like said before audio clock issue are fixed too.

[fedpul]

IMHO, I think what madshi is trying to do is let the options that really change IQ and remove the ones that are super power hungry for almost no visible difference, and I am not talking the difference you can see if you see pixel per pixel at 1cm from the screen, but at a normal view distance. It is like an audiophile that can claim that can hear differences between redbook standard and ultra HD tracks (means no offence, I am just being objective). Probably if madshi shows us two different settings telling which algos are in the presets he shows, and asking us to identify, we could fail or have 50/50 chances to be right. Like blind tests for audio files.

I repeat, this is just my opinion.

[leo91531]

Did you try software decoding? Had a similar issue with a R9 390 and 1080p 60fps H.264.

my LAV setting is HW Decoding.
http://imgur.com/t2LI2oa.png

when I ALT+TAB to desktop the rendering time will be down

sorry that GPU-Z Screen is too bright cause recording is not very good. You can look carefully that GPU load curve,it was not smooth and not stable.

[ikarad]

no not at all.

most content is 23p (24000/1001 hz) and 25/50 content will judder at 120 hz.
smoothmotion is very useful for 120 HZ screen and like said before audio clock issue are fixed too.
23.976p and not 23p.
I have 120hz and without smoothlmotion there is no judder.
I have tried smoothmotion and reclock to try to see any difference and I have never seen any difference.
In 60hz mode There is judder.


Madshi had said when I have posted this question many months ago that smoothmotoion was useless with 120hz monitor .

Maybe with 25/50 content but I don't have 25/50 content. I speak only about 23.976p content.

[huhn]

even windows and GPU driver are using 23p/23 hz for 24000/1001
and not to forget madVR...

but what ever.

playing a 23p file at 120 hz without reclock or smoothmotion will result in a frame repeate from time to time about all 42 sec. if you don't see that fine for you but it still happens.

something like that: https://abload.de/img/23pproblem.yyuz0.png

[Georgel]

Did you try software decoding? Had a similar issue with a R9 390 and 1080p 60fps H.264.

I have the same problem on GTX 1080.

There are some spikes to 200ms on decoding sometimes, but GPU usage is way under 100%, but it doesn't impact frame rate...



GTX 1080 can do everything at max, but on img upscaling luma / chroma / downscale at high on NGU with this version - but cooling is too noisy so it's not a profile to run for everyday enjoyment.

Settings for enjoyment ///

chroma / NG / high
downscling / JINC
UPSC / NGU / luma high, chroma / auto, downscale / medium - activate only if needed

refinement thin edges at 4.2 after every 2X step

I guess that my settings are not half as crazy as I made it seem throughout the time.

There is still a considerable amount of grinding / coil whine that is not there without NGU, but the image quality is way too good to skip.

Now it's the time I watch my poorly encoded Anime Music Videos and old Anime with a better image!

/This thing makes Fate anime series look like image quality is on steroids!

[JarrettH]

IMHO, I think what madshi is trying to do is let the options that really change IQ and remove the ones that are super power hungry for almost no visible difference, and I am not talking the difference you can see if you see pixel per pixel at 1cm from the screen, but at a normal view distance. It is like an audiophile that can claim that can hear differences between redbook standard and ultra HD tracks (means no offence, I am just being objective). Probably if madshi shows us two different settings telling which algos are in the presets he shows, and asking us to identify, we could fail or have 50/50 chances to be right. Like blind tests for audio files.

I repeat, this is just my opinion.

Yep. I mean when you get down to it, for instance with chroma upscaling, super-xbr brings it to a good quality level, but I hope you don't blink (and your nose must be touching the screen) when looking for improvements after :D

[aufkrawall]

I have the same problem on GTX 1080.

There are some spikes to 200ms on decoding sometimes, but GPU usage is way under 100%, but it doesn't impact frame rate...

You mostly have to stay below 90% GPU usage with madVR to avoid drops. I don't have any spikes with GTX 1070, a lot of load and DXVA2-cb.

[Telion]

madshi, anti-ringing checkbox in image downscaling is greyed out for me. Switching to chroma upscaling and back makes it active until I close the settings window.

The same goes for "image upscaling" when non-doublers are selected.

I would want to choose the upscaling algorithm to be used after image doubling for further upscaling to the target resolution. That algorithm may also be used for upscaling to lower scaling factors which don’t activate the selected image doubling algorithm. For instance, if NNEDI3 or NGU is selected in upscaling settings to be used, say, when scaling factor is 1.5x or bigger, Bicubic60 AR is used for upscaling to less than 1.5x scale factor whereas I would want to use Jinc there.

+1. Doublers shouldn't be mutually exclusive with non-doublers. I also vote for some kind of a custom/advanced checkbox for doublers which would switch their drop-down content from quality profiles to specific algorithms so their combination can be selected more freely.

[ikarad]

even windows and GPU driver are using 23p/23 hz for 24000/1001
and not to forget madVR...

but what ever.

playing a 23p file at 120 hz without reclock or smoothmotion will result in a frame repeate from time to time about all 42 sec. if you don't see that fine for you but it still happens.

something like that: https://abload.de/img/23pproblem.yyuz0.png

With reclock, I have frame repeated about all nearly 5 min.
I can have the same thing or even better (6.3 min for frame repeated) if I create custom resolution with 119.88hz (exact multiple of 23.976).

With 119.88 or 120hz or reclock, I have never seen the difference. Have you a recording to see the difference?

[Crimson Wolf]

I too have been using Catmull-Rom AR LL for downscaling after 720p to 1080p doubling, with previous madVR builds. Bicubic150 or SSIM1D look too sharp to my eyes in some cases.


I agree.

I would want to choose the upscaling algorithm to be used after image doubling for further upscaling to the target resolution. That algorithm may also be used for upscaling to lower scaling factors which don’t activate the selected image doubling algorithm. For instance, if NNEDI3 or NGU is selected in upscaling settings to be used, say, when scaling factor is 1.5x or bigger, Bicubic60 AR is used for upscaling to less than 1.5x scale factor whereas I would want to use Jinc there.

Correct me if I'm wrong, but can't you do that with profile groups now? Even with the new GUI? Just set srcWidth <= ??? "bicubic60ar + double scaling with NGU or NNEDI3" and srcWidth > "jinc" (since jinc disables doubling)?

I'm quite liking the new GUI so far.

[kalston]

Correct me if I'm wrong, but can't you do that with profile groups now? Even with the new GUI? Just set srcWidth <= ??? "bicubic60ar + double scaling with NGU or NNEDI3" and srcWidth > "jinc" (since jinc disables doubling)?

I'm quite liking the new GUI so far.

I think it's not too much to ask to be able to chose which upscaler to use after the doubling like we could before. I'm sure you can find a workaround in the meantime but wasn't this GUI update supposed to streamline things? :p

Or maybe he should just make JincAR the default instead of bicubic60, there are still some integrated GPUs that can choke on it (my Surface 3's with some content) but it should be fine for the vast majority of users.

[Telion]

madshi, since v0.91.2 I have a problem when NGU image upscaler uses Bicubic60 to process chroma. The picture is a mess of coloured rectangles and mostly freezed while I can see that the luma part is moving. Switching between 100% window and windowed FS modes or changing the size of a window, I noticed that the freezed image is based on the last one that was shown correctly in a 100% or below-doubling upscaled window. Even if I restart the player and go FS before opening the same video, the freezed image would be linked to a previous run of the player. If it is a new video, the freezed image would be based on its first frame. If I switch to NGU high and select NGU to also double the chroma there is no such problem, and when using non-doublers either. I can't test NNEDI3 on my ancient rig and it is on XP, so I suppose the problem is related to its old presentation path since no one has reported it yet.

Debug log & screenshot (http://wikisend.com/download/106092/report.7z)

[Neo-XP]

Correct me if I'm wrong, but can't you do that with profile groups now? Even with the new GUI? Just set srcWidth <= ??? "bicubic60ar + double scaling with NGU or NNEDI3" and srcWidth > "jinc" (since jinc disables doubling)?

I'm quite liking the new GUI so far.

I don't think there is a way with profiles to override the forced algorithms used after image doubling unfortunately.

For instance, to use Catmull-Rom AR LL for downscaling after 720p to 1080p doubling.

[Asmodian]

For instance, to use Catmull-Rom AR LL for downscaling after 720p to 1080p doubling.

I recommend that everyone who wants a particular option available post a screenshot showing the improvement with your content. madshi has always been responsive to suggestions when shown real benefits. :)

[Backflash]

no not at all.

most content is 23p (24000/1001 hz) and 25/50 content will judder at 120 hz.
smoothmotion is very useful for 120 HZ screen and like said before audio clock issue are fixed too.

It doesn't work. not at 120 or 144hz, it simply doesn't switch on itself unless you force it
it works only for very rare 30 fps on 144hz, it doesn't work on 120

[Backflash]

NGU still oversharpens noticeably on mid-quality SD content. I don't want to resort to enhancements to mess with the image, I'd like to use a clean image doubler instead. But by your argument, I'm already doing it right and NGU should do well. It's simply too harsh of an upscaler.

Can you post a screenshot please? Maybe i'm testing on wrong SD content, just curious.

Agreed. I just want to be able to use Lanczos3 AR again for luma/chroma upscaling and luma downscaling, after NGU. The new settings don't give me the same result as before. This is a regression for advanced users.


Please post comparisons, personally i'm simply curious what I can't see you can.

And you can actually use lanczos for upscaling luma/chroma but without doubling. But didn't it do the same job before?
ngu double->lanczos upscaling of doubled image if resolution was still too low pretty much negating ngu double?

[Schwartz]

Can you post a screenshot please? Maybe i'm testing on wrong SD content, just curious.

I tried, anyway. ;) NGU does have nice detail, and it's still fast. sXBR is softer overall and doesn't introduce any patterns even on higher sharpness.

Look at Deacon's hand or the windowblinds for example. There's a fuzzyness that you can spot in the cereal lettering. Look at the second house from the right, as well as the brick fence. NGU turns lines into jaggies. On the plus side it curbs ringing better than sXBR. These shots are without any further enhancements, just chroma, doubling and scaling.

Also when watching this SD stuff with compression artifacts in it, NGU makes these much more obvious in the moving picture, while sXBR tends to gloss over them. Stills don't really show this too well.

Anyway.. I don't expect you to agree with this. I like the soft-to-medium scalers the most so on that point I'm probably the odd man out.

https://s14.postimg.org/cjdrjc2fh/S1_NGU_High.png (https://postimg.org/image/cjdrjc2fh/)https://s14.postimg.org/ungs3z03x/S1_NGU_Low.png (https://postimg.org/image/ungs3z03x/)https://s14.postimg.org/688k2x171/S1_s_XBR_High.png (https://postimg.org/image/688k2x171/)https://s14.postimg.org/gjkwvkswd/S1_s_XBR_Low.png (https://postimg.org/image/gjkwvkswd/)https://s14.postimg.org/3tgoihky5/S2_NGU.png (https://postimg.org/image/3tgoihky5/)https://s14.postimg.org/5mjl6t64t/S2_s_XBR.png (https://postimg.org/image/5mjl6t64t/)

[QBhd]

I think NGU wins in all of the above images... I also detest softness, so that may be why I think NGU wins

QB

[Backflash]

Anyway.. I don't expect you to agree with this. I like the soft-to-medium scalers the most so on that point I'm probably the odd man out.


Yeah, I don't really understand that, I hate blur, I have enough of that in real life. I think you can lower those artifacts on NGU a bit with thin edges and sharpen edges 0.5(or even lower) with AR and optional AB. Image enchantments tab, they are not resource hungry.
LumaSharpen sometimes helps, but not with every source.
You probably tried that already, works for me, maybe will for someone else.

Oh and also do not forget about soften edges in refinement.

[Mistery 73]

I need your advice I have a nvidia 1070 GTX video card do you think it's worth having a GTX 1080 to make better use madVR?

[Betroz]

I need your advice I have a nvidia 1070 GTX video card do you think it's worth having a GTX 1080 to make better use madVR?

Upgrading from 1070 to 1080 is not worth it. Maybe upgrading to 1080Ti in january, or wait for Volta.

[Mistery 73]

thank you Betroz