madVR - high quality video renderer (GPU assisted)

[TheLion]

Originally Posted by Eyldebrandt

BTW, I have a question/request for future.
Is it technically possible to add in madVR a sort of internal supersampling ? I explain. With 1080p screen, you can't obtain the benefits of image doubling and SuperRes without some tricks.
If madVR could let the user set an internal resolution to reach by image doubling (+SuperRes etc..), with downscaling set up, the benefits of quality in HQ content could be very substantial.
It's just a thought, because it's currently feasible, but it's really not ergonomic.

"Internal Supersampling" is something I have suggested and discussed with Mathias in the past. With new options like SuperRes I would love to see this option especially for HQ 1080p content on 1080p displays. In my opinion this may prove to be a very beneficial approach for a "sharpening effect" by supersampling with SuperRes.

I hope Mathias sees value in it! I would appreciate it very much.

[TheLion]

Originally Posted by James Freeman

@ madshi, everyone.

Here is a great chroma sample: http://www.mediafire.com/?nyzrd9qt6nv9t41
*I removed all traces of the original name of the movie, but if it is still found somewhere, all rights reserved to the original creators.

Highly academic and certainly not comparable to real world content, but still my first choice for clearly showing the performance attributes of different chroma upsampling algorithms: https://drive.google.com/file/d/0B9J...ew?usp=sharing

With real content as well as with most other test patterns like e.g. the one linked by viewer (thanks for that!) differences are way too difficult to clearly spot in my experience.

@ madshi:
This is also one example where the new Bilateral option performs especially weak ("sharpness", high frequency detail extraction) imho - SuperRes improves it somewhat.

Interestingly the new super-xbr (especially without SuperRes) shows massive artifacts with this chart when the (new) antiringing filter is activated!

[James Freeman]

Originally Posted by TheLion

Highly academic and certainly not comparable to real world content, but still my first choice for clearly showing the performance attributes of different chroma upsampling algorithms: https://drive.google.com/file/d/0B9J...ew?usp=sharing

Wow thanks for that!

Lanczos 4 tap is a clear winner with this pattern, no visible Moire at all, nor brightness shift.
Bilateral again, is the weakest of them all, it simply destroys the chroma information.

[TheLion]

Originally Posted by James Freeman

Wow thanks for that!

Lanczos 4 tap is a clear winner with this pattern, no visible Moire at all, nor brightness shift.
Bilateral again, is the weakest of them all, it simply destroys the chroma information.

I agree. With this specific test pattern Lanczos8 AR was even better, practically the ultimate performing algorithm But this isn't the whole story...

[James Freeman]

This chroma test pattern also shows that LinearLight in downscaling is a MUST, to look anywhere like the fullsize image.
To me LL always looked a lot closer to the original in term of brightness when watching in a small window.

Try something funky:
Bilateral for chroma, and no LL in downscaling.
You can go no further from the source....

I settled on:
Upscaling: Lanczos4 + AR
Downscaling: Catmul-Rom + AR + LL
Chroma: Lanczos4 + AR

[nijiko]

Originally Posted by madshi

I'm not sure. I would have thought that it works. If it doesn't then please file a bug report on the madVR bug tracker. Thanks.

I don't know how to use your bug tracker.
If the var is after the symbol, the editor says it's invalid.
Such as "> srcWidth".

[Shiandow]

Originally Posted by TheLion

Highly academic and certainly not comparable to real world content, but still my first choice for clearly showing the performance attributes of different chroma upsampling algorithms: https://drive.google.com/file/d/0B9J...ew?usp=sharing

With real content as well as with most other test patterns like e.g. the one linked by viewer (thanks for that!) differences are way too difficult to clearly spot in my experience.

@ madshi:
This is also one example where the new Bilateral option performs especially weak ("sharpness", high frequency detail extraction) imho - SuperRes improves it somewhat.

Interestingly the new super-xbr (especially without SuperRes) shows massive artifacts with this chart when the (new) antiringing filter is activated!

I would highly advise against using that image to test the new algorithms. It doesn't seem to contain any luma information which messes up both the Bilateral algorithm and SuperChromaRes. Arguably the most accurate result would be a pure black image.

[madshi]

Originally Posted by viewer

It seems like no need Chroma SuperRes.

So which chroma upscaling algo do you prefer?

Originally Posted by aufkrawall

I can live without softness option with HQ downscaling enabled.
I think it makes more sense to be able to control strength instead of passes. I think it's a good compromise to have only one pass, but HQ downscaling enabled. Both for performance as for quality.
There are still huge differences between low strength and high strength with these settings.
So my vote goes for just keeping strength as an option, setting passes to 1, enabling HQ downscaling and setting softness to 0.
The strength option could be even more simplified by replacing the value box with templates like we have it with deband. Then it'd be just "SuperRes: low, medium, high".
Edit: But maybe it would also make sense to use 2 passes and instead lowering the strength a bit if one wants the SuperRes effect more distinct with the high setting.

Ok, thanks.

Originally Posted by aufkrawall

Well possible, the pink flowers also look darker with NNEDI3 64 chroma scaling.

Lower sharpness makes the pink even darker, like if it gets mixed more with green due to blur.

Try disabling the super-xbr AR filter.

Originally Posted by Eyldebrandt

if I upscale 1080p Blu-ray to 4K, regardless the algo in Double, low amount of passes in SuperRes will result with a lack of precision. So I have to multiple the passes, and with this state, strength > 0.5 will result to some pixels moving, and you have to renounce using any sharpen if needed. If you add some softness, you can adjust, but why add another process to maybe reach a targeted result you can win with 1 process in less ?
0.35 is a good median setting for clean regular skin, with fantastic precision, and the possibility to add some sharpening if I want/need.
To be simple, in 1080p to 4K, 5 passes @ 0.35 are far more accurates and faithfuls than 1 passe @ 1.0.
But, with 480p to 1080p, definitively, 1 passe @ 1.0 will be better than 5 @ 0.35.

So you like 5 passes @ 0.35 better than 1-2 @ 1.0. Is that visible in a screenshot? Or only in motion? If it's visible in a screenshot, can you make a comparison screenshot that shows the benefit of using more passes with lower strength settings?

Originally Posted by Eyldebrandt

BTW, I have a question/request for future.
Is it technically possible to add in madVR a sort of internal supersampling ? I explain. With 1080p screen, you can't obtain the benefits of image doubling and SuperRes without some tricks.
If madVR could let the user set an internal resolution to reach by image doubling (+SuperRes etc..), with downscaling set up, the benefits of quality in HQ content could be very substantial.
It's just a thought, because it's currently feasible, but it's really not ergonomic.

Originally Posted by TheLion

"Internal Supersampling" is something I have suggested and discussed with Mathias in the past. With new options like SuperRes I would love to see this option especially for HQ 1080p content on 1080p displays. In my opinion this may prove to be a very beneficial approach for a "sharpening effect" by supersampling with SuperRes.

I hope Mathias sees value in it! I would appreciate it very much.

It's technically possible. I'm not sure it will actually look better, though. Maybe I'll make a test build some time, but let's take it one step at a time. For now we're collecting feedback on SuperRes with 200% scaling factor.

Originally Posted by James Freeman

Here is a great chroma sample: http://www.mediafire.com/?nyzrd9qt6nv9t41
*I removed all traces of the original name of the movie, but if it is still found somewhere, all rights reserved to the original creators.

Thanks!

Originally Posted by tFWo

OK.

Here is an example with a ringing/blooming problem (and other artifacts with hq on). I'm using "refine the image only once" option now. It's an SD source upscaled with s-xbr50 quadrupling. Image is additionally zoomed in 2x in MPC-HC.

no superres
http://i.imgur.com/9R65zEz.png

2pass strength 0.50 softness 0.00 nohq
http://i.imgur.com/mAFZDxk.png

2pass strength 0.50 softness 0.00 hq
http://i.imgur.com/91RmoRJ.png

2pass strength 1.00 softness 0.00 nohq
http://i.imgur.com/hmREzhz.png

2pass strength 1.00 softness 0.00 nohq
http://i.imgur.com/LL7DAKs.png

and to compare with 0.88.13 and the settings i used before (refine the image after every ~2x upscaling step, 2pass, strength 0.50, sharpness -0.50, softness 0.15, AA 0.10, AR 0.10)
http://i.imgur.com/WDyf1Rc.png

I prefer the old Superres look and versatility .

Ok, thanks. Maybe something for Shiandow to look into?

Originally Posted by TheLion

Highly academic and certainly not comparable to real world content, but still my first choice for clearly showing the performance attributes of different chroma upsampling algorithms: https://drive.google.com/file/d/0B9J...ew?usp=sharing

With real content as well as with most other test patterns like e.g. the one linked by viewer (thanks for that!) differences are way too difficult to clearly spot in my experience.

@ madshi:
This is also one example where the new Bilateral option performs especially weak ("sharpness", high frequency detail extraction) imho - SuperRes improves it somewhat.

Interestingly the new super-xbr (especially without SuperRes) shows massive artifacts with this chart when the (new) antiringing filter is activated!

This test image/video has one purpose, and one purpose, only: To look good if the chroma upscaler rings a lot. The more the upscaler rings, the better this image looks. So I would say any algorithm which looks good with this image, should *not* be used in real life.

Ok, to be fair, I'm a bit overdramatic here. This image does have another benefit for chroma upscaling testing: Basically it allows you to check if an anti-ringing algorithm is too strict or not. A good anti-ringing algorithm should not make this image look too bad, just a bit worse than with the AR filter off. So this image clearly shows that the strict AR algorithm I'm currently using for super-xbr is really *very* strict. But on the positive side, it surpresses *all* ringing, while being very fast. I could implement the same AR algorithm which I'm using for super-xbr luma AR, but it would slow down the super-xbr chroma upscaling quite noticeably. And having a strict AR algorithm can sometimes be beneficial, as well (see the blue artifacts (or lack thereof) in viewer's chroma upscaling comparison).

Judging chroma upscaling algorithms is difficult. Judging AR quality is difficult, too. FWIW, Bilateral chroma upscaling and chroma SuperRes both look at the luma channel for guidance. And this specific test image/video has an empty luma channel. That's why it makes Bilateral and chroma SuperRes look very bad. That's not necessarily an indication of real world performance, though.

For a fair and objective analysis of chroma upscaling algorithms, multiple different test images/videos should be used. If you use only one image/video, you're getting only half the picture (or less).

That said, I'm mostly interested in SuperRes feedback atm.

[aufkrawall]

Originally Posted by madshi

Try disabling the super-xbr AR filter.

Yes, it's the AR filter which leads to the darkening.
But without, the ringing is very annoying. The ringing filter of Jinc doesn't have this issue.

Btw: Lanczos upscaling looks terrible with the cartoon sample. Ringing outta hell...

[madshi]

Oh well. As said above, I could implement a better AR algorithm for chroma super-xbr, but it would be expensive (performance wise).

[SecurityBunny]

Originally Posted by madshi

The problem is the following: As long as I allow D3D11 to use its default refresh rate, the presentation queues seem to fill nicely. But D3D11 is very reluctant to switch to some refresh rates sometimes. I've implemented a hack to work around the problem, which forces D3D11 to switch to our wanted refresh rate. But with some GPUs and some GPU drivers this seems to trigger this problem with queues not filling. Unfortunately I don't think there's much I can do about it. It seems that your D3D11 GPU drivers simply don't like outputting anything but 59.940Hz, for some reason. I'm not totally sure if it's D3D11 itself, or the GPU drivers which are at fault. Could be either, but I would guess it's probably the GPU drivers.

You could try creating custom resolution for 24Hz output. Or try using this tool:

http://www.monitortests.com/forum/Th...on-Utility-CRU

Maybe that helps, or maybe not. At this point the only thing I can offer is to either remove the hack I implemented, or to live with the consequences. If I remove the hack, you'll always get 59.940Hz, though, which is probably very far from what you want, too.

You can try some D3D11 games. Can any of them switch to anything other than 59.940Hz with your GPU? I doubt it...

Perhaps that may explain why present queue doesn't fill all the way initially when using 24hz, but it doesn't explain the bug. The bug occurs on all refresh rates as far as my testing goes.

The bug is when toggling to exclusive fullscreen, then windowed, then back to fullscreen again, the rendering and present queue completely drop and the playback stutters. This only occurs with D3D11 10bit. You need to pause the video and unpause it to get the queues to fill properly again. Using D3D9 10bit or D3D11 8bit, I can freely toggle in and out of exclusive fullscreen without queues not filling.

Plus the fact it takes longer going fullscreen with D3D11 10bit than compared to D3D9 10bit/D3D11 8bit.

[JarrettH]

I don't know if this is the best example, but no one seems to be posting film shots...

jinc chroma and luma
https://farm1.staticflickr.com/325/1...0025878e_o.png

xbr chroma jinc luma
https://farm1.staticflickr.com/405/1...4808c583_o.png

xbr doubling jinc chroma
https://farm1.staticflickr.com/421/1...7a8bb604_o.png

xbr doubling xbr chroma
https://farm1.staticflickr.com/554/1...6a825444_o.png

As you can see, it's difficult to tell apart chroma methods, even doubling

Any thoughts?

[huhn]

chroma it is just scaled to 4:4:4 the rest is done using image scaling. and chroma scaling has little effect anyway on such an image.

[leeperry]

TYVM for providing xbr-25, I didn't have time yet but I'll recompare it all again tomorrow. This said I was plenty happy with NEDI+SR in 88.13 and I'm afraid that the EE in xbr and NNEDI3 is just part of their design, I mean you want sharper edges you got them duh.......NEDI looks more natural and seems to be less agressively seeking edges, different methods for differents needs IMO. NEDI is perfect for 720@1080p, xbr50 for tiny videos but ideally I would like to find an in-between for untouched DVD's that require something sharp but not overly.

Originally Posted by aufkrawall

I think it makes more sense to be able to control strength instead of passes.

Totally, I don't use any kind of sharpening and if there has to be only one knob for SR please make it strength so we could all finetune it depending on our rig and personal taste.

[ryrynz]

Originally Posted by madshi

Why *higher* sharpness? I'd rather say lower sharpness, because Jinc is quite a bit softer than super-xbr 100.

Higher and lower would be good for comparison but my thoughts were if Jinc's ringing was benefiting it then higher sharpness for super-xbr would increase the ringing too, but if you say lower sharpness is a more fair comparison then sure.
I just figured aufkrawall had already determined 100 was the most suitable sharpness to compare Jinc against, but obviously not.

Originally Posted by viewer

It seems like no need Chroma SuperRes.

Any chance you can provide those as individual images?

[Warner306]

Originally Posted by madshi

Oh well. As said above, I could implement a better AR algorithm for chroma super-xbr, but it would be expensive (performance wise).

I think super-xbr150 + AR chroma is a small improvement over Jinc3 + AR. As long it does not incur a large performance hit, it is my choice for chroma upscaling for the future. The improved sharpness is apparent with real-world content.

[aufkrawall]

I personally rather achieve this with new SuperRes for chroma + Jinc AR.

[RyuzakiL]

Originally Posted by aufkrawall

I personally rather achieve this with new SuperRes for chroma + Jinc AR.

In regards to chroma upscaling (for low-res movies) which is better?

1. Jinc+AR+SuperRes (Default 2/1/0)

or

2. S-XBR-100+AR+ SuperRes (Default 2/1/0)

coz for the life of me i cannot discern any noticeable difference.

but if S-XBR is now much faster than Jinc then i think i will stick to number 2 config.

p.s. if I use S-XBR for Image Doubling does checking Interop Hack adds additional performance? or it is only applicable when using NNEDI3?

[Vyral]

@RyuzakiL

IMO and from what I've read on the thread, the principal difference is that sXBR is faster than Jinc. Also, you can't discern any noticeable difference because chroma upscaling has little impact on the image quality compared to image upscaling.

[aufkrawall]

With movies I hardly see any difference with any HQ chroma upscaling method so far. However, I could imagine that there are edge cases in which the ringing sensitivity can make a difference with low quality sources, but I doubt it would be critical.
So I assume it's not a bad idea to rely on super-xbr for chroma if you aren't watching cartoon content which isn't good quality in terms of ringing.