madVR - high quality video renderer (GPU assisted)

[Neo-XP]

I don't know why, but v0.91.1 is sharper than v0.91.4 with the same settings.

Here is v0.91.1 with NGU-med vs v0.91.4 with NGU-high, using 2x supersampling :

https://s11.postimg.org/b7hsk8gdb/old_med.png (https://postimg.org/image/b7hsk8gdb/) https://s4.postimg.org/thrntxvah/new_high.png (https://postimg.org/image/thrntxvah/)

Edit : I can see a difference in the stats displayed...

Edit2 : Maybe Lanczos3 AR for downscaling uses the strict (soft) setting in the new version ? @madshi ? Any way to change this ?

[MariaX9]

Chroma now is always at very high when i use luma very high in image upscaling, because let madVr decide and chroma very high are the same. I thought that was the reason for the gui change, cause ppl used chroma with ngu, or i'm wrong?

[omarank]

* redesigned "image upscaling" settings page once more
* super-xbr image doubling is available again
* fixed: downscaling anti-ringing checkbox was sometimes disabled
I hope you guys will like the new "image upscaling" settings page. I've compromised a little, gave you back some more control, but tried to keep it easy to understand for normal users.

I am happy with the new changes. There is one thing though: in case of NNEDI3 selection, chroma doubling at high uses NNEDI3 16 neurons, which I think is unnecessary. It can be replaced with Jinc for better performance. If you must offer NNEDI3 16 for chroma doubling, you can perhaps offer a medium setting as well for Jinc chroma doubling.

In case of NGU and suber-xbr, the available options are all perfect now. Thank you!

[CruNcher]

I don't know why, but v0.91.1 is sharper than v0.91.4 with the same settings.

Here is v0.91.1 with NGU-med vs v0.91.4 with NGU-high, using 2x supersampling :

https://s11.postimg.org/b7hsk8gdb/old_med.png (https://postimg.org/image/b7hsk8gdb/) https://s4.postimg.org/thrntxvah/new_high.png (https://postimg.org/image/thrntxvah/)

Edit : I can see a difference in the stats displayed...

Edit2 : Maybe Lanczos3 AR for downscaling uses the strict (soft) setting in the new version ? @madshi ? Any way to change this ?

Old seems indeed slightly sharper on some textures look at the finger and hand texture result alone :)

a difference of both would be interesting surely not worth the processing overhead in this specific noise/grain reconstruction example of yours though, this goes far away from most Anime flat frequency usage here, very nice.

8-10 ms is a rather huge difference for a inferior end result.

Great work you do here very different from most others :)

Btw you could see it in the above Human Portrait example already not only in this frame :)

The Perception Quality of the old is way nicer at less processing cost for your very specific test case, you should stay with it after a temporal evaluation shows the same results :)

But compared to this frame where the processing cost seems enourmous, im pretty sure you cant pinpoint the exact render cost frame accurate at all with the OSD, so this would need much more advanced testing ;)

PS: Your input is really needed keep it on

[hannes69]

That sounds pretty great!! I've put a LOT of work into NGU, so I'm happy to hear it pays off. JFMI: Could you maybe show a snooker screenshot comparing NGU vs whatever you were using before?

I have made some screenshots for you.
As I already mentioned, I don´t normally watch single pictures and spot for certain regions when magnifying and so on; I watch video content and go for the overall impression of the running video material.
For me the "wow-effect" of NGU is a good portion more emphasized in the running video, here for snooker especially when the balls are moving, the camera is moving and you´re seeing a player more zoomed.
JincAR and NNEDI64 are very similar to me, both look blurrier than each NGU to me, both have a natural look.
NGU low and med are the winners for me, both sharper than Jinc and NNEDI, both looking natural enough. NGU high and very high seem to be a little bit sharper than low and med but start to get an unnatural and processed look which is obvious the most when looking at zoomed people´s faces.
NGU med is the winner for me.

https://s15.postimg.org/frg8b3r5j/NGUlow.jpg (https://postimg.org/image/frg8b3r5j/)
NGUlow

https://s15.postimg.org/6xpbu046v/NGUmed.jpg (https://postimg.org/image/6xpbu046v/)
NGUmed

https://s15.postimg.org/ol74s7e47/NGUhigh.jpg (https://postimg.org/image/ol74s7e47/)
NGUhigh

https://s15.postimg.org/oc9k2a1br/NGUveryhigh.jpg (https://postimg.org/image/oc9k2a1br/)
NGUveryhigh

https://s15.postimg.org/5etxp0xmf/Jinc_AR.jpg (https://postimg.org/image/5etxp0xmf/)
JincAR

https://s15.postimg.org/quv92yn1z/NNEDI64.jpg (https://postimg.org/image/quv92yn1z/)
NNEDI64

[huhn]

So quad NNEDI3 & sxbr isn't possible anymore? :o

looks possible to me:
https://abload.de/img/q4ljss.png

[kalston]

Thanks for the new release madshi this is great now.

I usually find it hard to tell the difference between algos with most of the content I watch (on a 1440p display) but I'm noticing a real difference between NGU very high and NNEDI3 256+SR (yes, my 980 ti has no problem with either) for the few DVDs I watch (very excited about 3x NGU upscaling for this though since this would bring it very close to my native resolution!). I have noticed that SR brings NNEDI3 quite close to NGU for DVDs but still a bit softer and with some loss of detail. The big thing here is that NGU is a good 5ms faster despite doing 2x2x + SSIM downscale versus doing 2x+JincAR+SR for NNEDI3. I think that's exactly what you were hoping for. Have to say though I'm still impressed with what JincAR + SR can do this for this type of upscaling (yes it's softer but still good).

For 720p to 1440p, NGU is only better if I don't enable SR with NNEDI3. But again, it's much faster. NGU very high > 15ms, NNEDI3 256+SR > 25ms.

When watching high quality 1080p sources (blu-rays) the difference is invisible to me versus JincAR though, but at least I can do NGU doubling very high + downsampling (with NNEDI3 I had to drop the neurons quite a bit to play around with supersampling on blu-rays). I think in summer where I'm concerned about the heat in my room I'll just drop back to JincAR for this content. No matter how hard I try, can't spot the slightest difference.

I still have no opinion about chroma upscaling (as in, I can't see it), I used to have NNEDI3 / 32 neurons to bring it to the luma resolution (because I have the spare power) and then let madVR decide the rest. If I have performance issues it's the first place where I'll make compromises. Maybe I'll just NGU low for that now (a bit faster).

[thighhighs]

madshi
NGU image tripling is coming? I am waiting it for 720p -> 2160p and some sd -> full-hd upscaling.
Anyway. With new NGU image doubling logic (always if upscaling needed) i cant use v-high NGU luma for SD -> Full-HD upscaling, it resource intensive. High NGU quadrupling better vs v-high doubling + upscaling if source 720x400 -> 1920x1080 target?
And can you return\make image doubling settings (not "image doubling" section) for NGU and others (activate if it usefull, always... etc)?

[fedpul]

Thanks for the new version madshi, I found that it has advantages for new users and experts alike. I also would like to ask for an OSD feature. Could you add LL (Linear Light) in the OSD? I would like to know when it is being used or not.

[Telion]

That's really weird, especially because you seem to be the only one affected by this problem. Have you tried a different GPU driver version?
No, I have a very old and stable system and don't feel like messing around with it right now. But in v0.91.4 I've found out that the problem now appears only when in "algorithm to use after doubling"<--"upscaling" I choose Bicubic 60 AR. Every other available algorithm is OK, even just Bicubic 60, while chroma is still upscaled with Bicubic 60 AR always. So it's really weird, but if you won't be able to figure it out - it's not a big deal now.
madVR v0.91.4 released
* fixed: downscaling anti-ringing checkbox was sometimes disabled
This is fixed only in "image downscaling" and still not fixed in "image upscaling" when non-doubling algorithm is selected, as I noticed earlier.
I've dropped the option to let you choose with which exact scaling factor doubling gets activated. madVR now makes that decision itself. For NGU, doubling is activated as soon as any upscaling is needed, because NGU has some nice properties (e.g. removes ringing) which can be beneficial even when upscaling only by a very small factor. NNEDI3 and super-xbr are only activated if the scaling factor is 1.2x or higher.
I think this is a limitation that can affect weak GPUs, 1.2x is too low IMHO. And the decision logic for NGU is even more inconvenient. For example, if I choose NGU low for a video that upscales 2.4x it goes "NGU low > upscale", but for NGU med or higher it goes "NGU med > NGU low < downscale" and is too much for my card, while in previous versions I was able to just use "NGU med > upscale" in this situtaion. And going quadruple for a 2.4x factor is a bit off anyway. I understand that most users here have high-end GPUs and it's not a problem for them, but please don't left behind us resource-starved people and allow more flexibility in this aspect.
Of course you can use profiles to force any doubler to activate or deactivate in specific situations.
You mentioned earlier that you're aiming the profile system at advanced users while regular ones should be able to live without it. But now, stripping off convenient controls and limiting choices, you're inevitably pushing more and more people to use it. Isn't it a kind of contradictionary logic?

[Warner306]

What setting is used when upscaling after doubling is set to "let madVR decide?"

Can we consider this UI fixed for the coming weeks, or are you still planning on changing it?

[yok833]

Hello guys,
As always thank you Madshi for the new build!!
Is someone can explain my how to not use image doubling when I am watching a 1080p source on my 1080p tv?
Thanks!

[CruNcher]

@Neo-XP
Your second example with that enourmous overhead could be though interesting for anti aliasing edges the first though has a supreme sharpness reproduction you would want that for a uncompressed 4K result especialy

you also see that reflecting in the actuall jpg filesize

it did a good job in the whole chain reconstructing the quantization loses of the source encode :)

Especially on a 4K Display you see that difference, the blur is gone :)

That you saw that difference tells me you have very good trained eyes and a very good output device ;)

Hardly and of the ANIME guys could see that difference trained on their low frequency content ;)


So the first chain you chose is asolutely fine for your goal the second is just a pure waste of resources.

But the input must be already of a very good psy quality overall seing all that noise/grain preserved from the 35 mm master Scan :)

[Asmodian]

What setting is used when upscaling after doubling is set to "let madVR decide?"

Can we consider this UI fixed for the coming weeks, or are you still planning on changing it?

I think we should wait on any major updates, there is still room for change in the near future, especially if NGU direct tripling and quadrupling comes online. :)

Unless madshi thinks it is very unlikely? :D

Is someone can explain my how to not use image doubling when I am watching a 1080p source on my 1080p tv?
Thanks!

Don't check the "activate 200% super sampling" box.

[Warner306]

Is linear light necessary when downscaling after image doubling?

[Alamei]

Hello guys,
As always thank you Madshi for the new build!!
Is someone can explain my how to not use image doubling when I am watching a 1080p source on my 1080p tv?
Thanks!
I think you misunderstood one of madshi's earlier comments about when image doubling triggers automatically. The only time you would end up with image doubling on a matching source/output combo (i.e., 1080->1080 or 2160->2160) is if you explicitly check the supersampling option.

[Asmodian]

Is linear light necessary when downscaling after image doubling?

Not in my opinion, I like linear light when downscaling from the source resolution but not when downscaling to a resolution larger than the original source.

[fedpul]

Not in my opinion, I like linear light when downscaling from the source resolution but not when downscaling to a resolution larger than the original source.

That is the main reason that made me ask for LL in the OSD. Do you think it would be useful to have this info on OSD?

Like image<SSIM 1D 100 AR LL

I won't be using LL after doubling.

[aufkrawall]

I think I will leave soften edges set to 2 when upscaling <1080p to 1440p with NGU very high. It helps to prevent some edges being over-prominent and also helps a bit to make the contrast between edges and artifacts like ringing less obvious.

Btw: I'd vote for a warning with NGU + deringing like we have for SuperRes. It just looks terrible, not just with cartoons.

[fedpul]

madshi, I have noticed that you can't double SD with NGU like in 0.91.3, now if you choose NGU it just doubles twice. There are no options to choose from to make this not happen. Bug or feature? I know that in the future NGU will offer the option to just double or quadruple, maybe even triple the image resolution.

[StinDaWg]

madshi, I have noticed that you can't double SD with NGU like in 0.91.3, now if you choose NGU it just doubles twice. There are no options to choose from to make this not happen. Bug or feature? I know that in the future NGU will offer the option to just double or quadruple, maybe even triple the image resolution.
I went from being able to use NGU-High doubling plus upscaling, to only now NGU-Low with 480p60->1080p.

Doesn't look that great to be honest nor does it make much sense. :o

It looked better before with NGU-High doubling+Bicubic 60AR (since luma wasn't select-able), but there isn't any option to do this now. Quadrupling with low is not an option either so I can't tell what that looks like.

I totally understand wanting to limit the amount of options and making things more simple, but unfortunately this is killing the usability of weaker cards since there are so many fewer combinations of settings to tweak.

[Backflash]

Edit2 : Maybe Lanczos3 AR for downscaling uses the strict (soft) setting in the new version ? @madshi ? Any way to change this ?

what if you set it to "use image downscaling settings", there is such option exactly for that

[AngelGraves13]

madshi, is there a way to disable chroma doubling when using NNEDI3 or NGU? I'm going mostly from 1080p to 1440p, so it's doubling to 2160p and then back down to 1440p. Isn't the chroma scaler already going straight to 1440p and doubling it is a needless step adding overhead?

[kalston]

madshi, is there a way to disable chroma doubling when using NNEDI3 or NGU? I'm going mostly from 1080p to 1440p, so it's doubling to 2160p and then back down to 1440p. Isn't the chroma scaler already going straight to 1440p and doubling it is a needless step adding overhead?

For typical content (4.2.0) the chroma is half the res of the luma. So there should be one doubling step before the upscale to your native res. It's working as it should for me. What are your settings like?

[CarlosCaco]

No one miss sxbr ab 25? On very low quality movies maybe sxbr ab 25 than NGU...
Madshi it's possible to add a only luma supersampling option?
For performance reasons...

[Thunderbolt8]

don't you have it on tray?
just right click tray icon->edit madvr settingsI was asking for situations in which you are lying on the couch with a wireless keyboard/mouse and dealing with hdmi input lag of your tv so you would want to have to deal with all that as little as possible. it might look very small depending on the size and location of your 1st screen compared to where you arte situated at this moment.

[AngelGraves13]

For typical content (4.2.0) the chroma is half the res of the luma. So there should be one doubling step before the upscale to your native res. It's working as it should for me. What are your settings like?

Chroma: NGU high
luma and chroma doubling: NGU (high setting)

Jinc for upscaling and Jinc for downscaling.

About 20ms for 1080p content

[Neo-XP]

what if you set it to "use image downscaling settings", there is such option exactly for that

I didn't see that before ! Thanks !

The new UI is so confusing... it doesn't make sense for me to propose a few algorithms and an other option to select exactly the same algorithms with more possibilities.

Also, there is not such option for upscaling after doubling, and that is really the big problem of the new UI. You have to choose NGU OR another algorithm for luma and chroma upscaling, where you should be able to choose NGU AND another algorithm for them.

Maybe NGU (and other doublers) should be in another page (or another tab ? or separated by a horizontal line in the same page ?) so we can choose "use luma upscaling settings" and "use chroma upscaling settings", to upscale after doubling without any redundant choice in drop-down lists.

I think it is possible to make it simpler than it is now, while leaving every combination available and maybe even more than before. Going back to v0.91.1 for now, but I'm really looking forward to see what the next releases will offer in terms of UI and fine-tuning for users that don't have the latest GPUs.

NGU is so good that madshi could remove all other doublers now, they are obsolete. Prove me wrong :cool:
Testing is right now with "soften edges" at 2 and "add grain" at 3, the result is perfect in motion.

[madshi]

would it be possible to add a keyboard shortcut that opens the madvr settings tab at once in mpc-hc when playing a video? then youd dont have to right click --> filters --> madvr --> edit settings each time.
Yes, but you're the first one asking for this. Does anybody else need this? I'm not going to add features just for one user.

Could you please make the NGU+SR warning pop-up a one-timer?
Don't you remember that you already asked me for that? You know, if you ask me the same thing twice, that actually decreases the chances of being implemented.

Oh, and there used to be an option to only double for 1.2x and above, this was a nice feature IMO...
Please read the v0.91.4 announcement post. I've explicitly commented on this.

The ultimate question here is however how the performance/quality is in the end.

- Would NGU quadruple be faster then NGU+upscale?
- How does the quality compare, can a lower NGU when quadrupling match performance to double+upscale, at still better quality?
My original goal was to surprise everyone and create quadruplers which have almost the same speed as the doublers. But it didn't work out, sadly, image quality was subpar. So I had to make the quadruplers more complex. Currently my hope is that I might be able to make NGU direct quadrupling (at least "Low" + "Med" quality levels) roughly as fast as "NGU doubling + JincAR upscaling", while producing better image quality than doubling with NGU twice.

The "High" and "Very High" quality levels might end up somewhat slower than that, though.

Oh and chroma NGU "normal" doesn't actually do anything advanced IIRC? Might be nice to make it clear, it's Bicubic60AR doubling or so? So basically it's not possible to disable chroma doubling anymore as it was before?
I'm confused. You're saying you're getting Bicubic60AR chroma upscaling (which is *not* doubling), then you're saying you can't disable chroma doubling. That kind of contradicts each other.

I have made some screenshots for you.
As I already mentioned, I don´t normally watch single pictures and spot for certain regions when magnifying and so on; I watch video content and go for the overall impression of the running video material.
For me the "wow-effect" of NGU is a good portion more emphasized in the running video, here for snooker especially when the balls are moving, the camera is moving and you´re seeing a player more zoomed.
JincAR and NNEDI64 are very similar to me, both look blurrier than each NGU to me, both have a natural look.
NGU low and med are the winners for me, both sharper than Jinc and NNEDI, both looking natural enough. NGU high and very high seem to be a little bit sharper than low and med but start to get an unnatural and processed look which is obvious the most when looking at zoomed people´s faces.
NGU med is the winner for me.
Hmmmm... The cue/stick is the biggest different to me. Look at how aliased it is with JincAR!! The video does have a lot of mosquito noise, though. That is really ugly...

The NGU direct quadrupling algos should produce a cleaner and more natural looking result than doubling twice. Maybe that will make it possible for you to use NGU-High or VeryHigh.

For 720p to 1440p, NGU is only better if I don't enable SR with NNEDI3. But again, it's much faster. NGU very high > 15ms, NNEDI3 256+SR > 25ms.
Depending on the source, NGU should be noticeably better than NNEDI3+SR. However, to be fair, when using SR, you don't really need to use NNEDI3-256. For example NNEDI3-64 should look pretty much the same, thanks to SR. But in my tests, NNEDI3+SR can't really compete with NGU, if the source has a high quality (sharp and no or only few compression artifacts).

NGU image tripling is coming? I am waiting it for 720p -> 2160p and some sd -> full-hd upscaling.
That was the plan. I ran into some problems with it, though, so I can't say for sure right now if I can make a tripling NGU variant that is noticeably faster than quadrupling. It's possible tripling could end up with the same performance as quadrupling. If that's the case I probably won't bother creating tripling, because quadrupling + cheap downscaling will then probably offer the same or better image quality at similar speed. But it's all not decided yet.

in v0.91.4 I've found out that the problem now appears only when in "algorithm to use after doubling"<--"upscaling" I choose Bicubic 60 AR. Every other available algorithm is OK, even just Bicubic 60, while chroma is still upscaled with Bicubic 60 AR always.
That's actually helpful information! But I won't have time for madVR again until next weekend.

This is fixed only in "image downscaling" and still not fixed in "image upscaling" when non-doubling algorithm is selected, as I noticed earlier.
Ooops.

What setting is used when upscaling after doubling is set to "let madVR decide?"
Depends on the algorithm. E.g. with NGU-VeryHigh JincAR upscaling is used after doubling. With the lower NGU variants Bicubic60AR is used instead.

Can we consider this UI fixed for the coming weeks
I'm not sure what you mean?

Could you add LL (Linear Light) in the OSD?
Yes.

Btw: I'd vote for a warning with NGU + deringing like we have for SuperRes. It just looks terrible, not just with cartoons.
It's on my to do list, but it's not as straightforward as it seems. E.g. what happens if a user has a profile for situations where no scaling is needed, and if "deringing" is only active in this very profile? The settings dialog is not intelligent enough to understand that the profile can never be active at the same time as NGU, so it would complain in this situation, which wouldn't be a correct warning in this case.

In theory a similar problem could also occur with the current NGU + SuperRes warning, but it's somewhat less likely there.

is there a way to disable chroma doubling when using NNEDI3 or NGU? I'm going mostly from 1080p to 1440p, so it's doubling to 2160p and then back down to 1440p. Isn't the chroma scaler already going straight to 1440p and doubling it is a needless step adding overhead?
It's a complex topic. E.g. when using NGU, chroma doubling with Bicubic60AR is more or less "free", and scaling luma+chroma down at the same time is also almost "free". With NNEDI3 the situation is a bit different, though. You could try if the "trade quality" option "scale chroma separately, if it saves performance" makes a difference. It won't for NGU, but it might for NNEDI3, I'm not sure right now.

I don't know why, but v0.91.1 is sharper than v0.91.4 with the same settings.

[...]

Edit2 : Maybe Lanczos3 AR for downscaling uses the strict (soft) setting in the new version ? @madshi ? Any way to change this ?
Good catch! Yes, it's using the "strict" setting, which probably explains the difference in sharpness. Well, you have full control if you use the "use image downscaling settings" option. If you do that and pick Lanczos3AR "relaxed" there, does it bring the sharpness back to the old level?

Chroma now is always at very high when i use luma very high in image upscaling, because let madVr decide and chroma very high are the same. I thought that was the reason for the gui change, cause ppl used chroma with ngu, or i'm wrong?
Can you post a screenshot with the OSD (Ctrl+J) turned on, and a screenshot of your "image upscaling" settings page, so that I can fully understand what's going on there for you?

in case of NNEDI3 selection, chroma doubling at high uses NNEDI3 16 neurons, which I think is unnecessary. It can be replaced with Jinc for better performance. If you must offer NNEDI3 16 for chroma doubling, you can perhaps offer a medium setting as well for Jinc chroma doubling.
I don't think JincAR is much better than Bicubic60AR for chroma doubling, while NNEDI3-16 can be a bigger improvement in some rare situations. If you disagree, maybe you can post a screenshot which proves that JincAR is visibly better than Bicubic60AR, and similar in quality to NNEDI3-16?

With new NGU image doubling logic (always if upscaling needed) i cant use v-high NGU luma for SD -> Full-HD upscaling
if I choose NGU low for a video that upscales 2.4x it goes "NGU low > upscale", but for NGU med or higher it goes "NGU med > NGU low < downscale" and is too much for my card, while in previous versions I was able to just use "NGU med > upscale" in this situtaion.
madshi, I have noticed that you can't double SD with NGU like in 0.91.3, now if you choose NGU it just doubles twice.
I went from being able to use NGU-High doubling plus upscaling, to only now NGU-Low with 480p60->1080p.

Doesn't look that great to be honest nor does it make much sense. :o

It looked better before with NGU-High doubling+Bicubic 60AR (since luma wasn't select-able), but there isn't any option to do this now. Quadrupling with low is not an option either so I can't tell what that looks like.
Ok, I hear you. Would it help if you could simply disable NGU quadrupling altogether?

[madshi]

The new UI is so confusing...
I disagree. Try to forget what you're used to. Try to look at the settings dialog as if it were for the first time. The whole concept of resolution doublers is difficult to understand for new users. So now they can simply activate NGU and they can keep everything else as "let madVR decide". That's *very* easy for new users, and they don't have to understand the whole concept of resolution doubling *at all*.

At the same time, if you want more control, you can switch all those drop-down-boxes away from "let madVR decide" and you can achieve almost all (useful) settings combinations. And if you want to go back to "madshi recommended" settings, you can simply switch the drop-down-boxes to "let madVR decide".

IMHO it's a good compromise which makes things easy for new users, but at the same time gives you all the control you need. (Ok, maybe an option to disable NGU quadrupling is missing.)

[MariaX9]

Can you post a screenshot with the OSD (Ctrl+J) turned on, and a screenshot of your "image upscaling" settings page, so that I can fully understand what's going on there for you?



http://www.directupload.net/file/d/4559/hbbyy6p8_png.htm
http://www.directupload.net/file/d/4559/ii7kzmkg_png.htm

I tought, that chroma doubling isnt recommended or needed with ngu, with this settings, madvr chooses the chroma very high setting in image upscaling.

chroma let madvr decide = NGU med
chroma very high = NGU med

i was just wondering.

[oddball]

Madshi, do you know why an app called Smart Share for sharing content over my network causes MadVR and MPC-BE to crash? It gives me the error 'unrar.dll is missing' before actually crashing. Uninstalling it brings back normality. App link is below for you to test. I'm using Windows 10 and 32 bit MPC-BE.

http://www.lg.com/uk/support/smart-share

[fedpul]

Ok, I hear you. Would it help if you could simply disable NGU quadrupling altogether?

Thanks for LL on the OSD. It could help if we can simply disable NGU quadrupling, it could be just a checkbox stating "prevent NGU from auto quadrupling" or something like this.

Thanks for hearing. As always I really appreciate your work.

PS: I have an idea for fully automation of scaling algorithms that would like to share with you. I am trying to think as a developer even if I am not one. I will send you an email with details. It could be nice for 1.0 release or even further.

[Georgel]

@madshi

After running some tests, NGU is generally sharper than NNEDI, and works well - somewhat better. But on Anime especially, if there are straight lines, the sharpening can induce a bit of ringing, and some lines can get erased, same for some particles or shapes that don't have a solid outline.

This is for a direct comparation between algorithms, regardless of other algorithms engaged, as long as the only thing changed is the chroma upscaling algorithm and the image upscaling algorithm, it feels like the oversharpening adds a tiny bit of bright parts around dark edges and this feels bad for anime.

Seems that engaging edge thinning kinda solves the problem, but subjectively speaking - it does happen.

[sneaker_ger]

Madshi, do you know why an app called Smart Share for sharing content over my network causes MadVR and MPC-BE to crash? It gives me the error 'unrar.dll is missing' before actually crashing. Uninstalling it brings back normality. App link is below for you to test. I'm using Windows 10 and 32 bit MPC-BE.

http://www.lg.com/uk/support/smart-share
And the crash only happens when using madvr, not when using other renderers?

[Georgel]

I've made 3 examples.

The problem is as it follows: Both XBR and NNEDI are way smoother for Anime if resized from a low resultion. NGU, while sharper, has aliasing and ringing that is quite visible. The black lines come aliased, "in steps" compared.

Super XBR:

http://i.imgur.com/1V92UUp.png

NNEDI:

http://i.imgur.com/Syc5uWm.png

NGU:

http://i.imgur.com/pFFLaeb.png



EDIT:::

I post this in hope it is helpful for development! I'm very happy with madVR results, and NGU is cosiderably better for many things, but I watch anime more than 90% of the times I'm using madVR, and maybe this can be improved somehow..

[valdeski]

I don't know if this is related to madVR or not but I've only experienced it with NGU enabled so I'm gonna report it hoping someone can replicate this issue.

MPC-HC 1.7.10 64-bit + MadVR any version with NGU enabled, GTX 970 latest Nvidia driver 376.09, Windows 10 up to date, no insider preview testing.

Opening a directory in MPC-HC (File - Open Directory..) results in a BSOD that points to nvlddmkm.sys+d938

Sometimes it happens sometimes it doesn't, I've had it happen more often when opening consecutive directories as in open a directory and then immediately opening another directory, and only seems to happen with NGU enabled.
It could be the Nvidia driver 376.09 or NGU, could be any since both were released very close and it never happened to me with super-xbr or NNEDI3.

[jkauff]

I disagree. Try to forget what you're used to. Try to look at the settings dialog as if it were for the first time. The whole concept of resolution doublers is difficult to understand for new users. So now they can simply activate NGU and they can keep everything else as "let madVR decide". That's *very* easy for new users, and they don't have to understand the whole concept of resolution doubling *at all*.

At the same time, if you want more control, you can switch all those drop-down-boxes away from "let madVR decide" and you can achieve almost all (useful) settings combinations. And if you want to go back to "madshi recommended" settings, you can simply switch the drop-down-boxes to "let madVR decide".

IMHO it's a good compromise which makes things easy for new users, but at the same time gives you all the control you need. (Ok, maybe an option to disable NGU quadrupling is missing.)
I agree with you, madshi, and although I don't consider myself an "expert" user, I think the new UI works pretty well for both novice users and "experienced" users.

That being said, many people still have a lot of ripped DVDs in their collections, including some poor quality sources from the 30s, 40s, and 50s. I'm one of them. NNEDI3 still works much better for SD sources, so I'm not sure NGU should be the default choice for everything. Unless, of course, NGU improves for these sources over time.

I think any newbie-friendly UI should have a DVD/Standard Definition option and a High Definition option. Doesn't have to be NNEDI3, there are other ways to make SD look good, but you should give people the chance to see madVR work its magic to make SD sources look like HD sources.

I also think an OSD line that shows the percentage of GPU use would be very useful for everyone. The average user may not have GPU-Z, and in any case it can't be used in FSE mode.

[pose]

I've made 3 examples.
Any image enhancements or upscaling refinement on? By the way you might want to use soften edges and add grain along with NGU on SD content.

[Telion]

Yes, but you're the first one asking for this. Does anybody else need this? I'm not going to add features just for one user.
Yes, it would be nice to have a hotkey for opening madVR settings quickly. Especially while in FS when reaching the tray icon is not much easier than going through "filters --> madvr --> edit settings".
Ok, I hear you. Would it help if you could simply disable NGU quadrupling altogether?
And what if one wants to quadruple videos depending on resolution? The problem is not only with quadrupling but also with doubling activated on the 1.2x hardcoded threshold (or even without it as with NGU) - it is too aggressive for some situations. Going to settings every time or making profiles is so less convenient than just having a threshold selection list for doubling activation that would also enable quadrupling on a simple 2x by seleced value basis. So we can set a desired value according to one's GPU power or any other considerations and not on a "one size fits for all" ideology. Also, I think it would be nice and more logical not to have a hardcoded JincAR fallback if a doubler is selected and not activated, but to apply the one selected in "algorithm to use after doubling"<--"upscaling" except when "let madVR decide" is selected.

[Georgel]

Any image enhancements or upscaling refinement on? By the way you might want to use soften edges and add grain along with NGU on SD content.

Will test and post if the results are positive.

For enhancements, *only edge thinning and debanding, neither of which will cause aliasing. Both of them are more litely to solve aliasing and help with general image feeling, though edge thinning induces it's own wave'y typical artifact if the factor is too high. Compared to the original, seems like the numbers I'm using are the sweet spot between edge thinning being too aggressive and not enough.

[clubby]

First let me say, I've used MPC-BE w/ madVR for years with excellent results. I use the 3D capable version in my projector room. I recently purchased an LG OLED 4K 3D TV and I have some questions/problems maybe somebody could help with.

I don't know if it's specifically a madVR problem, but I set my monitor resolution to 1920x1080 (3D). In madVR I have everything enabled for 3D, set the resolution to 1080p23 in madVR. When I start a 3D mkv movie, my TV switches into 3D mode. So far, so good. I tried three movies and I get similar behavior all three times (this is a fresh install of this setup and a brand new TV). The movie will play flawlessly for about 7-10 minutes and then the computer just hangs. It hangs with a perfect picture on the screen, but nothing I can do will allow it to exit. I end up having to do a hard reset of the computer.

I made two changes. I changed madVR to go into fullscreen exclusive mode rather than windowed. I also changed chroma>catmull rom to lanczos3. I then started another 3D movie and so far it's been playing for 32 minutes.

Do either of these things sound like changes that could have made a difference? I typically don't fiddle with things too much. This particular piece of software has always worked really well for me "right out of the box."

I will say this, I hope I can get this working without a lot of hiccups because the 3D image on this OLED 4K TV is absolutely stunning.

Any help would be appreciated. I have used this particular HTPC for many years with no trouble until this. It's particularly overbuilt for the application. 8 core AMD black processor, R7 280 video card, 16GB RAM, SSD.

Here's a picture of my stats running now.

http://i.imgur.com/7xnD2AH.jpg?2

[Asmodian]

http://www.directupload.net/file/d/4559/hbbyy6p8_png.htm
http://www.directupload.net/file/d/4559/ii7kzmkg_png.htm

I tought, that chroma doubling isnt recommended or needed with ngu, with this settings, madvr chooses the chroma very high setting in image upscaling.

chroma let madvr decide = NGU med
chroma very high = NGU med

i was just wondering.

That is the expected behavior, veryHigh is sort of like a placebo option so enabling chroma doubling at a low level at the same time makes sense. If using veryHigh over high is worth it then chroma doubling probably is too. :)

Chroma doubling isn't totally useless, it is simply less useful than increasing luma quality, to a point. If I could run NGU veryHigh for luma but only without chroma doubling I would, but NGU med for chroma is pretty easy relative to NGU veryHigh for luma so it is likely that someone who could run NGU veryHigh for luma would have the performance for both.

For me when doing 1080p -> 4K: NGU veryHigh for luma takes an extra 15.4ms while NGU med for chroma only takes an extra 3.0ms (compared to bicubic60AR).

[andybkma]

I think this is a limitation that can affect weak GPUs, 1.2x is too low IMHO. And the decision logic for NGU is even more inconvenient. For example, if I choose NGU low for a video that upscales 2.4x it goes "NGU low > upscale", but for NGU med or higher it goes "NGU med > NGU low < downscale" and is too much for my card, while in previous versions I was able to just use "NGU med > upscale" in this situtaion. And going quadruple for a 2.4x factor is a bit off anyway. I understand that most users here have high-end GPUs and it's not a problem for them, but please don't left behind us resource-starved people and allow more flexibility in this aspect.



madshi, I echo this sentiment... I use mVR on a laptop with a 650M and obviously can't upgrade the video card... and I also can't do any doubling at all so please leave any options in mVR that us low end video card users can still use...

[ryrynz]

I've made 3 examples.

The problem is as it follows: Both XBR and NNEDI are way smoother for Anime if resized from a low resultion.

We established this the same day it was released.

[Asmodian]

We established this the same day it was released.

This makes sense too, as a neural net trained awhile ago, on older video content, NNEDI3 does a great job with most SD material. NNEDI3 for SD, NGU for HD. :)

NGU seems to make an "HD version", which replicates all the flaws too well. Any video that looked like an SD anime when scaled to 50% would have some very obvious flaws. :(

[ryrynz]

This makes sense too, as a neural net trained awhile ago, on older video content

I suspect madshi wants to replace NNEDI3 so we might see another version of NGU aimed at lower res content.

[burfadel]

NGU:

http://i.imgur.com/pFFLaeb.png

[QUOTE=clubby;1788460]
Here's a picture of my stats running now.

http://i.imgur.com/7xnD2AH.jpg?2

All three screenshots by Georgel (only NGU shown above as no point showing the other), and the one by Clubby show a present queue that doesn't seem to be working.

Clubby has it set to 14 and Georgel 8, but in both cases they are running at only 1! I guess that's why there are dropped frames. I both cases the other queues are all perfectly filled.


The movie will play flawlessly for about 7-10 minutes and then the computer just hangs. It hangs with a perfect picture on the screen, but nothing I can do will allow it to exit. I end up having to do a hard reset of the computer.

I've come across that numerous times as well. People have reported various issues sporadically in this thread, I seem to be hit by most of them! I've finally worked out how to play it flawlessly with the right settings, the downside being that it is D3D11 windowed mode (but full screen) rather than D3D11 exclusive.

@Georgel
How come D3D 11 fullscreen windowed rather than D3D 11 fullscreen exclusive?

[ryrynz]

Please don't quote such large images in the thread, thanks.

[Sunspark]

I'm wondering something.. surely I can't be the only one that notices the aspect ratio differences when upscaling resolutions that aren't perfect 16:9s.. many of you now have high resolution monitors, mine is 2048x1152.

Two video files tonight, both anime:

1772x952 as reported in the OSD.
The default aspect ratio selected is 1.86 and it scaled to 0,26,2048,1126.
Changing it manually to 1.85 scales it to 0,22,2048,1129.

1280x688 as reported in the OSD.
The default aspect ratio selected is 1.86 and it scaled to 0,26,2048,1127.
Changing it manually to 1.85 scales it to 0,22,2048,1129.

I think the 1.85 setting looked a tiny bit better than the default. Circles are the best thing to look at when judging this.

Same deal with this video..

1920x808 as reported in the OSD.
The default aspect ratio selected is 21:9 and it scaled to 0,145,2048,1007.
Changing it manually to 2.35 scales it to 0,140,2048,1011.

I think the 2.35 setting looked a tiny bit better.

Basically I am seeking understanding of why the "default" calculation is picking what it does, rather than what I think it should be picking instead based on what I see for proportions.