madVR - high quality video renderer (GPU assisted)

[aufkrawall]

Originally Posted by madshi

I don't really know how to improve the deringing filter much further. It is what it is. At some point I might try to use the NGU algorithm framework to create a better deringing filter, but I'm not sure how slow that would be. Since NGU itself already derings pretty well I'm not sure how much need there still is for a separate deringing filter, when upscaling with NGU?

I was hoping that with your own scaler, deringing filter and upscaler could work more "coordinated" to prevent it turning the "reduced ringing" into "aliased ringing".
I replied to huhn regarding the anime example with MPEG2 artifacts:
http://forum.doom9.org/showthread.ph...98#post1786198

NNEDI3 64 + deringing:


NGU high:


The amount of ringing gives NGU quite some headaches.
That "aliased ringing" effect with NGU + deringing is very well visible with the other American Dad cartoon example.

[thighhighs]

super-xbr 100 doubling + SuperRes 3: 28ms
NGU-low doubling: 37ms

I use NGU without SuperRes. Doubling luma and chroma in both cases here. Image upscaling Jinc.
Another settings

NGU-low double luma only, image upscaling C-R + AR: 18ms. It fast.

[madshi]

Originally Posted by XTrojan

Used Superres for both, is NGU bad with superres or did i miss something?

As far as I can see, NGU produces results which are so good that SuperRes doesn't know how to further improve them. So using SuperRes with NGU simply doesn't bring any benefit. It may even harm image quality. It's better (for image quality and speed!) to use NGU on its own, without SuperRes.

[JarrettH]

Is it necessary for madvr to be 50 mb?

[Shiandow]

Originally Posted by madshi

Interesting. Why is SSIM's detail enhancement better in gamma light, if I may ask?

Linear light might be more physically accurate, but if you want to measure detail then gamma light is closer to how humans perceive detail. In theory something like CIELAB would be even better, but the difference are pretty small so I'm not sure if that will matter that much.

Obviously you run into problem because, paradoxically, humans estimate brightness as if the image was in linear light, which means some parts of SSIM should be done in linear light, while others should be done in gamma light (or Lab or similar), and I haven't entirely figured out which is which.

[madshi]

Originally Posted by thighhighs

NGU-low double luma only, image uscaling C-R + AR: 18ms. It fast.

And which has overall better image quality now? This setting? Or your super-xbr + SuperRes setting?

Originally Posted by aufkrawall

I was hoping that with your own scaler, deringing filter and upscaler could work more "coordinated" to prevent it turning the "reduced ringing" into "aliased ringing".

Sorry, but no. I don't know how to coordinate them.

[aufkrawall]

Originally Posted by JarrettH

Is it necessary for madvr to be 50 mb?

Seems like the archives are not much compressed anymore. I'd recommend 7z with ultra compression, archive is only 16MB of size then.

[cork_OS]

Originally Posted by madshi

As far as I can see, NGU produces results which are so good that SuperRes doesn't know how to further improve them. So using SuperRes with NGU simply doesn't bring any benefit. It may even harm image quality. It's better (for image quality and speed!) to use NGU on its own, without SuperRes.

Thank you, I just had a question about SR effect on NGU.
So maybe NGU should just ignore "SR on" mark?

[madshi]

Originally Posted by JarrettH

Is it necessary for madvr to be 50 mb?

I've all my pixel shaders precompiled in a million different versions. That consumes a lot of space. But it helps keeping madVR startup times small.

Originally Posted by Shiandow

Linear light might be more physically accurate, but if you want to measure detail then gamma light is closer to how humans perceive detail. In theory something like CIELAB would be even better, but the difference are pretty small so I'm not sure if that will matter that much.

Obviously you run into problem because, paradoxically, humans estimate brightness as if the image was in linear light, which means some parts of SSIM should be done in linear light, while others should be done in gamma light (or Lab or similar), and I haven't entirely figured out which is which.

Ouch. I wonder if getting this exactly right would noticeably improve image quality. I rather guess that the visible difference might be small. But it's a wild uneducated guess.

Originally Posted by cork_OS

So maybe NGU should just ignore "SR on" mark?

I've already thought about that. But leeperry would run amok if I did that...

[JarrettH]

Originally Posted by madshi

i've already thought about that. But leeperry would run amok if i did that...

:d :d

[heiseikiseki]

Originally Posted by madshi

madVR v0.90.1 released

http://madshi.net/madVR.zip

* added NGU "low" quality algorithm
* "soften edges" and "add grain" now allow smaller adjustment steps
* fixed: install batch didn't work
* fixed: on some PCs NGU didn't activate
* fixed: on some PCs keyboard shortcuts didn't work
* fixed: NGU grain was static
* "ShowRenderSteps" hack no longer slows down NGU as much as before
* "ShowRenderSteps" hack only slows down stuff as long as the OSD is visible

I hope that OSD and NGU now work for everyone (again)?

Please let me know what you think about the NGU "low" variant. With my AMD 7750 it's almost twice as fast as "medium", and it's now faster for me than Jinc/super-xbr + SuperRes 1. However, it depends on which image upscaling algorithm you've selected, because Jinc/super-xbr double both luma and chroma at the same time, while if you tell NGU to only double luma, chroma needs to be doubled separately. So if you use an expensive algorithm for chroma doubling, NGU-Low may be slower again than Jinc/super-xbr + SuperRes. So for best speed vs quality ratio I'd suggest to combine NGU-Low for luma doubling and to choose e.g. CatromAR for "image upscaling".

Everything works here.
Thanks so much!

And NGU looks amazing for Anime!

for gtx 860m (optimus)
720p->1440p in NGU-low (chroma and doubling) for 35ms which is near to jinc ar or super-xbr.

[madshi]

Originally Posted by heiseikiseki

720p->1440p in NGU-low (chroma and doubling) for 35ms which is near to jinc ar or super-xbr.

Is the doubling just for luma or for both luma and chroma? I'd suggest to only use luma doubling. And in image upscaling settings try CatromAR.

[Shiandow]

Originally Posted by madshi

Ouch. I wonder if getting this exactly right would noticeably improve image quality. I rather guess that the visible difference might be small. But it's a wild uneducated guess.

Yeah I wonder that as well... FWIW I'm pretty sure downscaling in gamma light causes problems, so either running SSIM in linear light magically causes no problems whatsoever, or there's still room for improvement.

Long story short, linear light SSIM requires more testing.

[madshi]

Originally Posted by Shiandow

either running SSIM in linear light magically causes no problems whatsoever, or there's still room for improvement.

Long story short, linear light SSIM requires more testing.

For now it seems to work well enough for me. When switching between gamma and linear light, the biggest differences I can see seem to be caused by the downscaling operation itself, not by the SSIM processing. Not 100% sure about that, though.

[aufkrawall]

Originally Posted by heiseikiseki

And NGU looks amazing for Anime!

Not just anime, everything that's not utterly broken.

I can use NGU low for 1080p -> 4k -> WQHD btw. even with deband and NGU low chroma scaler (but not doubler). But SSIM is too much, bicubic works.
Edit: Nope, I can't. Queues get empty.

[madshi]

One thing that would be worth investigating:

Does SSIM downscaling produce any visible benefits over e.g. Bicubic150, when downscaling an image that was doubled before?

The reason I'm wondering is that one key benefit of SSIM is reproducing texture detail nicely. But when doubling an image, there's not much texture detail in the doubled image. So I think Bicubic150 might work just as well as SSIM in that situation. It's quite possible that SSIM only really helps when directly downscaling a high-res source. Anybody up for testing this in depth?

[thighhighs]

Originally Posted by madshi

And which has overall better image quality now? This setting? Or your super-xbr + SuperRes setting?

Super-XBR doubling + SuperRes 3 not better, but more slow vs NGU-low luma\C-R AR (or lanczos3 AR, ~18ms too) in my small test.

[jerryleungwh]

I am using a 960M GPU and NGU-low for image doubling is too much for my GPU when I try to upscale 720p->1080p even when I'm just using bicubic75 for both chroma upscaling and image downscaling and is causing a lot of frame drops, what would be a recommendation for me to use the algorithm for image doubling?

And I have been trying to use NGU-med for chroma upscaling as that's the best my GPU can handle, so how does it compare with NNEDI32 SR1(which is what I have been using before the update) for chroma upscaling?

[Q-the-STORM]

win10 x64 + GTX 780 Ti SC
NGU works fine here, didn't have hotkey issues... madVR v0.90.0 seemed to start up very slowly (2-3 seconds instead of starting instantly), but v0.90.1 is starting instantly again, so that's back to normal...

When playing 1080p videos on a 2160p screen, using NGU doubling on chroma and luma, there should only be NGU-med showing up in the OSD, but mine is showing:
chroma > NGU-med
luma > NGU-med
chroma > lanczos3 AR


lanczos3 with AR is set in image upscaling, but since this is just 1920x1080 to 3840x2160 doubling, why is lanczos showing up in the OSD?

[madshi]

Originally Posted by jerryleungwh

I am using a 960M GPU and NGU-low for image doubling is too much for my GPU when I try to upscale 720p->1080p even when I'm just using bicubic75 for both chroma upscaling and image downscaling and is causing a lot of frame drops, what would be a recommendation for me to use the algorithm for image doubling?

Which algo for "image upscaling"?