madVR - high quality video renderer (GPU assisted)

[aufkrawall]

Thanks for that sample, leeperry.
I'm seeing too much aliasing with SuperRes at the balcony railing, even with strength 1 and radius 1.0.
Even NNEDI3 64 doesn't look exactly great here, 256 neurons really improve the image a lot. However, 64 neurons are much more reasonable and still look better than super-xbr (e.g. regarding the antennas).
AS 0.2 UR still makes the image a little sharper without really destroying anything with NNEDI3 64. That's exactly what I meant with "set & forget".

I scaled it to 1080p and compared it with the original, madshi.

[Thunderbolt8]

Im wondering in how far it makes sense to compare different upscaling/downscaling algorythms when the comparison pic has been downscaled/upscaled before. couldnt there be potential interference from the downscaling/upscaling process which has been done before the upscaling/downscaling process? perhaps the outcome would be different in some cases if the first step wasnt done?

so doesnt it make more sense to look at a picture which has only either been upscaled or downscaled and not both?

[Akeno]

Originally Posted by Thunderbolt8

so doesnt it make more sense to look at a picture which has only either been upscaled or downscaled and not both?

If we do that, we get back into subjectivity. What looks best to one will differ from person to person and that is a conversation that won't reach a definitive conclusion.

[ryrynz]

Originally Posted by har3inger

This may be GPU/driver specific. The only way to test for yourself is to take two screenshots of the same frame, one with DXVA2 native, one without, and subtract them from each other in photoshop or something.

Tested on my Intel HD 4000 with both madVR and MPDN and there's no difference.

[JarrettH]

Love the comparisons. I guess I'll hold out until I can run SuperRes on top of super-xbr doubling :P

[ryrynz]

Originally Posted by har3inger

This may be GPU/driver specific. The only way to test for yourself is to take two screenshots of the same frame, one with DXVA2 native, one without, and subtract them from each other in photoshop or something.

On my setup, if you compare DXVA2 native and DXVA2-cb, there are definite differences if you subtract the screenshots from each other that are greater than the difference you get from the different dithering pattern. All of these differences are barely visible, as they are under 5/255 luminosity in their intensity.

I'm on an ATI HD 8870 (R9 M270X) paired to an intel HD4000. Catalyst 15.6 beta

Originally Posted by ryrynz

Tested on my Intel HD 4000 with both madVR and MPDN and there's no difference.

Madshi, looks like there's an issue here. On my 750 Ti there are considerable differences between running LAV in copy-back mode and native DXVA2.

http://screenshotcomparison.com/comparison/136044

[nevcairiel]

Originally Posted by ryrynz

Madshi, looks like there's an issue here. On my 750 Ti there are considerable differences between running LAV in copy-back mode and native DXVA2.

http://screenshotcomparison.com/comparison/136044

Its a known issue, one that cannot be fixed properly (ie. without copy-back), which is why you should use copy-back.
madVR used to have its own copy-back function, but it was removed because LAV is just more efficient at it.

It seems to mostly affect NVIDIA, which is the good news, since copy-back is very efficient on nvidia anyway.

[ryrynz]

Originally Posted by nevcairiel

Its a known issue, one that cannot be fixed properly (ie. without copy-back), which is why you should use copy-back.
madVR used to have its own copy-back function, but it was removed because LAV is just more efficient at it.

It seems to mostly affect NVIDIA, which is the good news, since copy-back is very efficient on nvidia anyway.

Thanks, wish I had known this earlier. Not sure why this has really only been made apparent now, it's something that should be in the OP IMO considering the quality difference.

[nevcairiel]

Originally Posted by ryrynz

Thanks, wish I had known this earlier. Not sure why this has really only been made apparent now, it's something that should be in the OP IMO considering the quality difference.

Before evaluating the quality, you should also make sure that all trade quality options are off that interact with DXVA in any way. I don't think its usually *that* bad, but maybe I remembered wrong.

[ryrynz]

Originally Posted by nevcairiel

Before evaluating the quality, you should also make sure that all trade quality options are off that interact with DXVA in any way. I don't think its usually *that* bad, but maybe I remembered wrong.

All are unticked.

[chros]

Originally Posted by nevcairiel

Its a known issue, one that cannot be fixed properly (ie. without copy-back), which is why you should use copy-back.
madVR used to have its own copy-back function, but it was removed because LAV is just more efficient at it.

It seems to mostly affect NVIDIA, which is the good news, since copy-back is very efficient on nvidia anyway.

Thanks for the explanation!
Those test was done using DXVA native and copyback. What about CUVID? (I've been still using it for a long time.)

[huhn]

CUVID is just the room heating version of copyback.

[ryrynz]

Yeah I was wondering today when eventually it'll be removed.

[Thunderbolt8]

Originally Posted by Akeno

If we do that, we get back into subjectivity.

why, in how far is judging this different than judging it from an image which has been up- and downscaled? how can you only be sure the result you see is from the algorythm you want to look at and has not been influenced by the other one?

[leeperry]

Originally Posted by Akeno

I'm still under the impression that HQ is closer to the ground truth than LQ.

Fair enough, I'm sure madshi will explain me how wrong I am using 500% magnified close-ups but at the end of the day I kinda rest my case that SR is in dire need of at least 3 separate knobs for number of passes, strength and LQ/HQ.

Originally Posted by aufkrawall

256 neurons really improve the image a lot. However, 64 neurons are much more reasonable and still look better than super-xbr (e.g. regarding the antennas).

I meant to try 256 neurons NNEDI3 for chroma+luma but mVR .20 was refusing to scale the picture to 1080p for some reason, I seriously wish all this upscaling madness could be used on still pictures too coz it's either a black frames feast or a total refusal to show anything. All picture viewers provide craptastic upscaling algorithms.

Oh BTW I had a gamut mapping PS script engaged in PotP so yeah the ground truth is the one I labelled as is and not the untouched capture.

[aufkrawall]

Originally Posted by leeperry

I meant to try 256 neurons NNEDI3 for chroma+luma but mVR .20 was refusing to scale the picture to 1080p for some reason

Any idea why? I had doubling activated for >= factor 1.5 (default) and it worked just normally.

Btw: Is it technically possible to just sharpen chroma to fight blur of CSS?

[Ver Greeneyes]

Originally Posted by leeperry

I seriously wish all this upscaling madness could be used on still pictures too coz it's either a black frames feast or a total refusal to show anything.

I've had success loading images in MPC-HC, though it's not ideal for saving the result since I think madVR uses slightly different logic for that. I'm pretty sure the upscaling works just fine though.

[digitech]

Originally Posted by nevcairiel

Its a known issue, one that cannot be fixed properly (ie. without copy-back), which is why you should use copy-back.
madVR used to have its own copy-back function, but it was removed because LAV is just more efficient at it.

It seems to mostly affect NVIDIA, which is the good news, since copy-back is very efficient on nvidia anyway.

Which do you think is more efficcient? Im on nvidia 760 and still using Cuvid in Lav Filters, is copy back a better alternative?

[aufkrawall]

It should be in no way inferior. You probably even save energy with mpeg4-asp when NOT using Cuvid and instead software.
nevcairiel has no control over Cuvid since it's a blackbox, the DXVA2cb implementation is by himself.

[Akeno]

Originally Posted by Thunderbolt8

why, in how far is judging this different than judging it from an image which has been up- and downscaled? how can you only be sure the result you see is from the algorythm you want to look at and has not been influenced by the other one?

I'm not sure what you're trying to get at here. The nature of scaling involves throwing away details when downscaling and preserving what details still exist when upscaling. Unless you had an analog source captured at two different optical zoom levels, this is the best we can do. Most digital sources we receive are going to be upscaled or downscaled using algorithms.