madVR - high quality video renderer (GPU assisted)

[Q-the-STORM]

Originally Posted by madshi

The feedback I've received from several industry insiders has been quite different. For example, a couple weeks ago a film maker (advertising business, not Hollywood) asked me to make my HDR -> SDR conversion algorithm available for him, because he wants to grade only in HDR and then get a good SDR master by simply converting the HDR graded master to SDR with my algorithm. He says he's tried every HDR -> SDR algorithm on the market and mine produced by far the best results.

Any chance your algo will ever be available as a avisynth or vapoursynth plugin?

[Cinemancave]

Originally Posted by madshi

That sounds kinda weird. I'm not sure if I want to make this part of the GUI. Maybe it could be some undocumented hack option or so.

An undocumented hack would be fine. Otherwise we JVC owners would have to stop updating MadVR in order to keep this (weird) functionality.

A bit more info:When a JVC projector receives the HDR metadata-flag, it locks the projector into an "HDR"-mode that JVC (for some strange reason) has disabled the automatic iris off, which leads to much poorer black levels in low-APL scenes. The only solution so far has been to buy the HDFury Linker (199$) that has a feature to strip HDR-metadata. Hence, the projector still receives HDR but isn't locked into the "HDR"mode, which brings back the dynamic iris. This has a huge impact on black levels, and is currently the go-to solution for watching HDR on these projectors.

It would be awesome if MadVR could do this so that we can get the same results without having to involve the HDFury Linker. A simple hack would be welcome

[madshi]

Originally Posted by Neo-XP

Because it look better with good sources, "soften edges" (at 1 or 2, not more) only soften the most sharpen edges produced by "NGU Sharp", but "NGU Standard" applies a whole different process to the image and everything looks softer.

Oh well. How about we drop NGU Soft, and instead add a new preset in between NGU Sharp and NGU Standard? Would that allow us to remove "soften edges"? Do you have a good sample image where the different between the "whole different process" is extremely obvious?

Originally Posted by Neo-XP

What about NGU AA/Standard for chroma doubling in replacement or in addition to NGU Sharp ? It does not make sense that the "high" quality setting (NGU Sharp low) is worse than the "normal" one (Bicubic60 AR).

You can't judge these things with just one sample, especially not if the difference isn't really all that big. With really HQ sources NGU Sharp chroma doubling looks clearly better than Bicubic60 AR. E.g. try the well known Clown image.

Originally Posted by Neo-XP

Do you have to develop these NGU variants specifically for chroma doubling or could you just add them ?

I don't have to develop them speficially for chroma, but I also can't "just add them", because the chroma doubling option does not list any algorithms at all. I would have to change the logic of the chroma doubling option to list specific algorithms instead of "low, medium, high" quality.

Originally Posted by Q-the-STORM

Any chance your algo will ever be available as a avisynth or vapoursynth plugin?

Actually, there will be a new product available pretty soon, making most of the madVR algos available to avi/vapoursynth users. There will be a "pro" version of this new product and a free version. The free version will have some limitions, which will hopefully not be too dramatic, but motivate real "pro" users/companies to buy the pro version. This is all I can say about this for now. More details to be announced later.

[cork_OS]

Originally Posted by madshi

Chroma subsampling (4:2:2, 4:2:0) is always factor 2x. So what purpose would NGU chroma quadrupling have? Or do you mean when we start with 4:4:4, you want both luma and chroma to be directly quadrupled? That's already done today.

No, I meant something like this (for 720p->1080p):

Chroma > NGU Standard (low) 4x
Luma > NGU Standard (high)
Image < Bicubic150 AR

[madshi]

Originally Posted by cork_OS

No, I meant something like this (for 720p->1080p):

Chroma > NGU Standard (low) 4x
Luma > NGU Standard (high)
Image < Bicubic150 AR

I see. Yes, in theory that would be possible. However, in the long run the plan is to create an NGU chroma upscaling algo which makes use of the higher res luma channel. Since the luma channel is always max 2x as large as the chroma channel, this kind of algo will have to be 2x, too. This new kind of algo will hopefully beat all other algos, so I hope it will be the default algo at some point in the future. Because of this reason I'm not too fond of the idea of spending a lot of time now to implement a combination of chroma quadrupling + image doubling, although it is a nice concept.

[Selur]

there will be a new product available pretty soon, making most of the madVR algos available to avi/vapoursynth users.

Nice!

[Neo-XP]

Originally Posted by madshi

Oh well. How about we drop NGU Soft, and instead add a new preset in between NGU Sharp and NGU Standard? Would that allow us to remove "soften edges"? Do you have a good sample image where the different between the "whole different process" is extremely obvious?

But it will take you a lot of time to develop another NGU variant just for the sake of removing "soften edges", and there is already too many NGU variants.
So if you "need" to remove it, remove it, just don't waste your time on another NGU variant for this.

You can take this image to test with an upscaling to UHD : https://s15.postimg.org/ro3n1k6qz/original.png
Or here if you don't want to waste time :

NGU Sharp : http://www.mediafire.com/file/aj27g3.../NGU_Sharp.png
NGU Sharp SE2 : http://www.mediafire.com/file/d04aw9..._Sharp_SE2.png
NGU Standard : http://www.mediafire.com/file/wacasf...U_Standard.png

Originally Posted by madshi

You can't judge these things with just one sample, especially not if the difference isn't really all that big. With really HQ sources NGU Sharp chroma doubling looks clearly better than Bicubic60 AR. E.g. try the well known Clown image.

The Clown image is not representative of the content I use, which is mostly FHD HQ films for UHD upscaling.
On already sharp content, NGU Sharp (low) for chroma doubling is too sharp and quite nasty (see http://screenshotcomparison.com/comparison/206580 and http://screenshotcomparison.com/comparison/206596).

Originally Posted by madshi

I don't have to develop them speficially for chroma, but I also can't "just add them", because the chroma doubling option does not list any algorithms at all. I would have to change the logic of the chroma doubling option to list specific algorithms instead of "low, medium, high" quality.

Maybe the chroma doubling drop-down list should also lists algorithms with their names, preceded by the quality.

For instance, if I choose "high" for luma doubling, I would have this for chroma doubling :

- let madVR decide
- low (Bicubic60 AR)
- normal (Jinc AR)
- high (NGU AA low)
- high (NGU Standard low)
- high (NGU Sharp low)
- very high (NGU AA medium)
- very high (NGU Standard medium)
- very high (NGU Sharp medium)

or any other combinations of algorithms that you find adequate to use...
At the moment, new users do not even know which algorithm is used for chroma doubling without displaying the stats.

For fun, I tested Bicubic60 AR against Jinc AR : http://screenshotcomparison.com/comparison/206712
It is not difficult to see the red gradient lines/blocks with Bicubic60 AR, while Jinc AR is smoother and has the highest quality.

Originally Posted by madshi

Actually, there will be a new product available pretty soon, making most of the madVR algos available to avi/vapoursynth users. There will be a "pro" version of this new product and a free version. The free version will have some limitions, which will hopefully not be too dramatic, but motivate real "pro" users/companies to buy the pro version. This is all I can say about this for now. More details to be announced later.

If there is one day a "pro" version of madVR with less limitations, all algorithms available and more options, I'm in !
Maybe it is your goal to remove/limit things now from madVR in order propose them back in a "pro" version, so the actual version will remain free ?

[cork_OS]

Originally Posted by madshi

Now I need your FEEDBACK, once again:

1) Do you like the new "NGU Standard" and "NGU Soft" variants? Personally, I like "NGU Standard" a lot. But I think "NGU Soft" is too soft to be useful, and I'd like to remove it again. But since I spent a lot of time creating it, I thought I'd let you test it, at least.

2) Does "NGU Standard" work better for (non-aliased) low quality content now? Or do we still have to use NGU AA for that? Ideally I hope that we can now use "NGU Sharp" for high quality soures, "NGU Standard" for low quality (but aliasing free) sources, and "NGU AntiAlias" only for aliased sources. Is that reasonable? Or is "NGU AA" still the best option for any low quality sources, aliasing free or not?

3) I want to remove the "Soften Edges" option, because "NGU Standard" was made to replace it. Any objections?

4) With all the changes in this build, I feel that "double twice" is no longer needed. "direct quadrupling" should now always be identical or superior in quality and much faster. So I would like to remove "double twice". Any objections?

1) I like NGU Standard very much. It provides pretty natural look, quite sharp in the same time and without excessive line thinning. [EDIT: Agreed that NGU Soft looks unnecessary.] The only benefit of NGU Soft over NGU AA is removing of dark halo in some cases.

2) Should notice that "low quality content" includes content with very different quality flaws. For example, low quality sources may contain staircase noise, which is less amplified by NGU AA. In other hand, NGU Standard is immune to dark halos. However, in general NGU AA looks more suitable at least for very low-quality content, it brings uniform and natural-looking picture.

3,4) No sir!

Originally Posted by madshi

However, in the long run the plan is to create an NGU chroma upscaling algo which makes use of the higher res luma channel. <...> This new kind of algo will hopefully beat all other algos, so I hope it will be the default algo at some point in the future.

Reconstruction at the price of NGU would have been wonderful.

[Ver Greeneyes]

Originally Posted by madshi

Fair enough, for low quality content. But how about high quality content? Do you prefer NGU Sharp or NGU Standard?

I watch so little high quality content that I feel weird answering this - but on the parts of low quality content that don't look compressed I like the crispness of NGU Sharp.

Originally Posted by madshi

NGU AA is as sharp as it's ever going to be, I believe. Of course it's possible to post-sharpen it, but I don't think doing that will be part of the algorithm. I mean it might be possible to include artificial sharpening as part of the algo, but it probably wouldn't look any better than doing NGU AA first, and then apply post-sharpening.

Alright, I might mess around with post-sharpening then. I prefer using the built-in qualities of algorithms because it limits the amount of options I have to explore (and also reduces the overhead).

[leeperry]

Originally Posted by madshi

1) Do you like the new "NGU Standard" and "NGU Soft" variants? Personally, I like "NGU Standard" a lot. But I think "NGU Soft" is too soft to be useful, and I'd like to remove it again. But since I spent a lot of time creating it, I thought I'd let you test it, at least.

2) Does "NGU Standard" work better for (non-aliased) low quality content now? Or do we still have to use NGU AA for that? Ideally I hope that we can now use "NGU Sharp" for high quality soures, "NGU Standard" for low quality (but aliasing free) sources, and "NGU AntiAlias" only for aliased sources. Is that reasonable? Or is "NGU AA" still the best option for any low quality sources, aliasing free or not?

Hard to tell coz NGU AA can't do direct quad, would that be complicated to implement?

Originally Posted by madshi

4) With all the changes in this build, I feel that "double twice" is no longer needed. "direct quadrupling" should now always be identical or superior in quality and much faster. So I would like to remove "double twice". Any objections?

I think you got a point, but I'd love to see quad NGU AA before answering all your inquiries if any possible please.

It's really too bad you removed SR AB especially when NGU bloats quite a lot in combination with SR, I currently exclusively use SSIM downscaling(1D for HD, 2D for SD as I got GPU cycles to spare) and use SSIM's AB with great results

For 720p@1080p, NGU Sharp wins by a long shot IME.

[Q-the-STORM]

Originally Posted by madshi

Actually, there will be a new product available pretty soon, making most of the madVR algos available to avi/vapoursynth users.

Sounds great.

Originally Posted by Neo-XP

If there is one day a "pro" version of madVR with less limitations, all algorithms available and more options, I'm in !
Maybe it is your goal to remove/limit things now from madVR in order propose them back in a "pro" version, so the actual version will remain free ?

He's not talking about madVR, he's talking about a new product which makes madVRs algorithms available for avisynth and vapoursynth...

[Neo-XP]

Yes I know...

[madshi]

Originally Posted by Selur

Nice!

Originally Posted by Neo-XP

Maybe the chroma doubling drop-down list should also lists algorithms with their names, preceded by the quality.

For instance, if I choose "high" for luma doubling, I would have this for chroma doubling :

- let madVR decide
- low (Bicubic60 AR)
- normal (Jinc AR)
- high (NGU AA low)
- high (NGU Standard low)
- high (NGU Sharp low)
- very high (NGU AA medium)
- very high (NGU Standard medium)
- very high (NGU Sharp medium)

or any other combinations of algorithms that you find adequate to use...
At the moment, new users do not even know which algorithm is used for chroma doubling without displaying the stats.

What's the point of adding "low", "medium" and "high", if I list multiple different options for each? The whole point of the way the dropdown box works right now is to protect users from making crazy decisions, which the history of this thread shows is a real problem. I may switch to using medium instead of low quality for chroma doubling, though.

Originally Posted by Neo-XP

For fun, I tested Bicubic60 AR against Jinc AR : http://screenshotcomparison.com/comparison/206712
It is not difficult to see the red gradient lines/blocks with Bicubic60 AR, while Jinc AR is smoother and has the highest quality.

If you're honest, IMHO you would probably have to admit that the difference is incredibly small, and this at something like 5000% zoom, and with a test image which should be ideal at showing the difference. Don't you agree? With 99.99% other images, you wouldn't see any difference at all even at 500000% zoom. Would you really want to spend precious GPU resources on using Jinc AR for such a tiny quality difference at 5000% zoom?

Originally Posted by cork_OS

1) I like NGU Standard very much. It provides pretty natural look, quite sharp in the same time and without excessive line thinning. [EDIT: Agreed that NGU Soft looks unnecessary.] The only benefit of NGU Soft over NGU AA is removing of dark halo in some cases.

2) Should notice that "low quality content" includes content with very different quality flaws. For example, low quality sources may contain staircase noise, which is less amplified by NGU AA. In other hand, NGU Standard is immune to dark halos. However, in general NGU AA looks more suitable at least for very low-quality content, it brings uniform and natural-looking picture.

3,4) No sir!

Ok, thanks for the feedback!

Originally Posted by Ver Greeneyes

I watch so little high quality content that I feel weird answering this - but on the parts of low quality content that don't look compressed I like the crispness of NGU Sharp.

K, thx.

Originally Posted by leeperry

It's really too bad you removed SR AB especially when NGU bloats quite a lot in combination with SR, I currently exclusively use SSIM downscaling(1D for HD, 2D for SD as I got GPU cycles to spare) and use SSIM's AB with great results

NGU doesn't bloat. SuperRes reintroduces some bloating, but that's not NGU's fault. Too be honest I'm not too worried about it, because I think you're probably the only person on the planet who combines NGU with SuperRes.

[Neo-XP]

Originally Posted by madshi

What's the point of adding "low", "medium" and "high", if I list multiple different options for each? The whole point of the way the dropdown box works right now is to protect users from making crazy decisions, which the history of this thread shows is a real problem. I may switch to using medium instead of low quality for chroma doubling, though.

The point is to have the choice of the algorithm, while indicating the level of quality of each one so users don't make crazy decisions
If they don't know what they are doing, they should stick with "let madVR decide" anyway. That is why this option exists, right ?
Switching to NGU Sharp (medium) will not solve the problem, as it will still be too sharp. The only solution is other NGU variants (like you did for chroma upscaling).

Originally Posted by madshi

If you're honest, IMHO you would probably have to admit that the difference is incredibly small, and this at something like 5000% zoom, and with a test image which should be ideal at showing the difference. Don't you agree? With 99.99% other images, you wouldn't see any difference at all even at 500000% zoom. Would you really want to spend precious GPU resources on using Jinc AR for such a tiny quality difference at 5000% zoom?

Yes, I agree and admit that, but I think that you minimize too much the impact of chroma doubling in image quality.
I was not planning to use Jinc anyway if there is a good NGU alternative to NGU Sharp for chroma doubling, and Jinc has all the chances to become totally useless (performance and quality wise) with NGU Standard and NGU AA.

[leeperry]

Originally Posted by madshi

NGU doesn't bloat. SuperRes reintroduces some bloating, but that's not NGU's fault.

Which was the whole point of SR AB to begin with, I have yet to comprehend why this had to go

[Telion]

madshi, what about the usage of sharper relaxed AR instead of strict in downscaling after doubling? You said you could reconsider this and apply it more granular depending on the sharpness of a doubling algorithm, as on weak GPUs it adds an excessive load without visible benefits and thus sometimes prevents use of a doubling algorithm with better quality. Can you please apply strict AR at least on NGU AA and maybe also on super-xbr?

[direxx]

Hi,

various people stated that HDR does only work when using DXVA2 with copy-back on AMD Polaris. I couldn't figure out how to change the DXVA2 mode in madVR. I double-checked every single madVR option in the settings dialog, but could not find anything related to DXVA2/copy-back (I suspect it should be under rendering).

So how do I enable this mode in madVR?


Thanks, direxx

[steakhutzeee]

Ok so, in the past i solved the "flickering" in the black parts of the screen during videos, the letterbox black, changing my monitor settings from 60 Hertz to 75.
But why i still have this sort of horizontal semitransparent flickering only in the video portion of the screen now? If i don't use madvr the flicker goes away. Maybe it's a madvr setting i set not properly?

Just to be clear, letterbox black bars are not flickered, solved by changing Hertz frequence. But i still have it in the video portion "between black bars".

[Budtz]

Originally Posted by Budtz

No settings sadly fixes it for me

I did manage to get it somewhat stable with glitches and fremedrops only about every 15 minutes. before the creators update there none of these issues

threads are starting to pop op around the web with ppl having the same problems

I take that back. disabeling FSE fixes the issue. the system behaves a bit different but its fine.

I have another problem however. i get glitches if i use dxva in lav filters. what could couse this? cuvid is fine

[pose]

madshi, i tested new NGU variants with low res low quality anime (direct quadruple very high vs AA 2x doubling very high) and must say that i still like NGU AA the most. NGU Standard is much better than NGU Sharp for low quality anime as it has much less "oil painting" effect and less aliasing but has some "flat" look to it compared to NGU AA. NGU AA has visibly less focus and sharpness compared to NGU Standard but has less aliasing and looks more pleasing and more uniform to me. So far Standard looks like a good alternative to NGU AA for those that want something sharper but not NGU Sharp level of sharp and with slightly thinner lines. I dont really like NGU Soft as it has less focus than NGU AA, thins lines too much and has more aliasing.