madVR - high quality video renderer (GPU assisted)

[webs0r]

Barnahadnagy, blindbox already answered this. I'll just explain in different words in case it helps.

Game = don't know when next frame will come (e.g. could be 5ms later, or 60ms)
Video = know exactly the frame spacing (e.g. always 41.6ms for 24fps)

So for video, just match the monitor refresh to the video frame rate (or multiple). Gsync is useless in this case.

A case where it might be useful is if that monitor doesn't support a particular refresh rate. But then you run into that other problem, where if we drove it via software, the software is unlikely to be able to have it hit e.g. 48 Hz exactly. Games just present the next frame when its ready, it doesn't need to worry about "hang on wait until exactly 41.6ms from the last frame, then present". Computers aren't set up to time this precisely that well.

We'd really need those monitors to be able to switch to any arbitrary refresh rate - that would be more useful instead.

[madshi]

Originally Posted by toniash

" Superres is skipped in this situation if the overall upscaling factor is smaller than 1.125x"

Why?
SuperRes can be very useful as a sharpener even on not upscaled images.

The original purpose of SuperRes was to enhance/improve the image quality when doing upscaling. One integral part of the algorithm compares the upscaled image to the original image. Doing this doesn't make too much sense if the upscaling factor is very small.

But yes, SuperRes has some algorithms in it which "do something" even if you don't upscale at all. So yes, it could be an option to run SuperRes even if no scaling is performed at all, or even when downscaling. But this is really outside of what SuperRes was originally written for. I do wonder if running dedicated sharpening algorithms wouldn't produce better results when not scaling. Anyway, for now it is as it is. Once we've dumbed all the settings down to low/medium/high, we can discuss whether it might make sense to allow SuperRes also when not upscaling, but let's first reduce all those options, to make image quality comparisons easier.

Originally Posted by toniash

Another question: How processing changes when SuperRes is "applied first"?

It doesn't change much. This option only defines in which order FineSharp, LumaSharpen and SuperRes are applied, if you have enabled more than one of those 3 options in "upscaling refinement". That's all.

[Hyllian]

Originally Posted by madshi

If you could change that to (or dual license it as) MIT or BSD license, that would be most awesome! Of course I would give you proper credit in my readme file (and later in some credits window, if I ever add one, not sure yet). If you find MIT/BSD too generous, LGPL would also work, but if I could choose, I'd *strongly* prefer MIT/BSD.

There you have it --> https://github.com/libretro/common-s.../xbr/super-xbr

MIT license from now on! I was already willing to do it for some time.

[sandro08]

Hello everyone

I am new to the forum and I am the last few years
I ask your help for madVR passage 10 bit
I made some screenshots for you to see my parametrage
my graphics card is a nvidia 970gtx and my diffuser is a w1070 benq I think to be consistent after the explanations given here
My problem is the following

I can not seem to have the 10-bit option in the control panel nvidia, I can only access 8 bits
I plug in hdmi







excuse me if the question was asked already

friendly

[huhn]

your screen doesn't support 10 bit.

maybe a limitation of your HDMI cable but very unlikely.

[sandro08]

Originally Posted by huhn

your screen doesn't support 10 bit.

maybe a limitation of your HDMI cable but very unlikely.

ok though when I did the test ca looks ok
cacle my hdmi is 1.3

thank you

[yok833]

Originally Posted by sandro08

ok though when I did the test ca looks ok
cacle my hdmi is 1.3

thank you

I have the same issue with my Panasonic plasma St50 which is quite recent too... It can only support 12bit with 23hz but no option available with 60hz (I tried with an hdmi cable highspeed 2.0)... So I guess my TV cannot support it...

[Della]

Originally Posted by huhn

your screen doesn't support 10 bit.

maybe a limitation of your HDMI cable but very unlikely.

Just curious, how can you tell that from the information he provided?
Simply because the CP doesn't present the option?
Thanks

[Shiandow]

Originally Posted by madshi

In some image areas ringing is beneficial, in others not. I think we would get better results by using your algo once, then sharpening the result in a clever way, then run your algo again, then sharpen it again. The sharpening should have a similar effect to running jinc. Of course the sharpening would need to have some anti-ringing incorporated, too. A combination of LimitedSharpenFaster (AviSynth script) and FineSharp should work well.

Might be interesting to combine super-xBR with SuperRes (or one of it's variations), I suspect that those two will combine especially well.

[Hyllian]

Originally Posted by Shiandow

Might be interesting to combine super-xBR with SuperRes (or one of it's variations), I suspect that those two will combine especially well.

super-xbr has a internal param that let you control the level of edge detection. In those images it's set to 1.0. If I had set it to 0.0, it would try to find edge even in grass and other textures. I think it has the same objective of SuperRes, right?

[Shiandow]

Originally Posted by Hyllian

super-xbr has a internal param that let you control the level of edge detection. In those images it's set to 1.0. If I had set it to 0.0, it would try to find edge even in grass and other textures. I think it has the same objective of SuperRes, right?

Not exactly, the objective of SuperRes is to upscale an image such that downscaling it will give the original image. It then imposes some constraints on the image to make this process well defined and numerically stable.

So far one of the biggest problems has been that this process doesn't necessarily remove aliasing. If done well it doesn't add any, but removing it still seems a bit difficult. However it is pretty good at recovering detail. So the idea is that if you combine it with an algorithm which is good at removing aliasing, you get an image with low aliasing and good sharpness.

[Hyllian]

Originally Posted by Shiandow

Not exactly, the objective of SuperRes is to upscale an image such that downscaling it will give the original image. It then imposes some constraints on the image to make this process well defined and numerically stable.

Interesting. Though I think it's a strange objective, because I don't know if it can guarantee you'll get a good result. Depending on the downscale method, a nearest neightbor upscaled image would be a perfect result for SuperRes.

[madshi]

Originally Posted by Hyllian

Interesting. Though I think it's a strange objective, because I don't know if it can guarantee you'll get a good result. Depending on the downscale method, a nearest neightbor upscaled image would be a perfect result for SuperRes.

It does sound a bit strange, but as you say it depends a lot on the algorithm used for downscaling, and on the algorithm used to upscale the difference/error comparison result back to the upscaled image size. If you use linear sampling for downscaling and nearest neighbor for upscaling the difference/error, then probably the perfect SuperRes result would be nearest neighbor. But if you use better downscaling/error upscaling algorithms, that is no longer true. E.g. imagine you downscale the NNEDI3 (or your) image back to the original image size, using a decent downscaling algorithm. The difference to the original image will be very *very* small. So you take that very small difference, scale it back up to the upscaled image and apply it to the upscaled image. It does not really result in nearest neighbor at all, if you upscale the error/difference with a good algorithm. E.g. using Jinc upscaling, the upscaled difference should be pretty much free of aliasing. The only situation where the difference gets bigger is if the original upscaling algorithm made some really bad decisions somewhere. E.g. this whole concept works quite well to remove some of the directional artifacts of the NEDI algorithm (not NNEDI3, but NEDI).

That said, I'm not totally convinced yet that SuperRes is better than "simple" good sharpening, when starting with a good artifact-free upscaling algorithm. Still need to do some more investigation into the matter. SuperRes is a relatively "young" algorithm.

@Shiandow, what I've found is that the SuperRes downscaling/difference/upscaling sequence has somewhat similar effects to an unsharp mask. The downscaling/upscaling is somewhat comparable to the gaussian blur in the unsharp mask. Which means that the upscaled image is sharpened by SuperRes, even if you set sharpness to 0.0, and also gets some ringing. It might make sense to include some anti-ringing surpression right when applying the difference to the upscaled image, instead of relying on the extra anti-ringing pass. It's usually better for image quality to avoid adding ringing instead of trying to remove it afterwards. Just a thought, though...

[madshi]

P.S: @Hyllian, thank you very much for relicensing your algorithm & shaders under MIT, that's very generous! I'll try to implement your shaders in madVR during the weekend.

I suppose your images in this thread have been created with passes 0-1, is that correct? Or have you used passes 0-2 for those?

[Hyllian]

Originally Posted by madshi

P.S: @Hyllian, thank you very much for relicensing your algorithm & shaders under MIT, that's very generous! I'll try to implement your shaders in madVR during the weekend.

I suppose your images in this thread have been created with passes 0-1, is that correct? Or have you used passes 0-2 for those?

They were created with the three passes, though the third pass is optional.

I think the third pass improves a bit the IQ, in exchange of power, of course. I don't know if it's worth. Anyway, it's optional and can be skipped if power is a problem.

As you'll see, my implementation has much room for optimization. I calc the luminance inside all shaders and it could be done in a first pass (see NEDI shaders, for example) only.

The first pass must work at 2x scale factor (the same way NEDI shaders), and the second and third passes work at 1x scale factor. Use nearest sampling on all texture lookups, as it doesn't need precise texture sample locations.

You'll see too that my blendings aren't done in linear light. I didn't test this algorithm in linear light.

[ryrynz]

Originally Posted by madshi

P.S: @Hyllian, thank you very much for relicensing your algorithm & shaders under MIT, that's very generous! I'll try to implement your shaders in madVR during the weekend.

Awesome Hyllian, looking forward to the test build.

[MysteryX]

Originally Posted by Hyllian

Ok, for your wishes --> http://i.imgur.com/C1xv6Pr.png

That's using the shaders two times. The anti-ringing is ON two times!

Personally I find the image unnatural. Too "cartoonish" for my taste. But then that probably works for Mario Bros

Originally Posted by madshi

That said, I'm not totally convinced yet that SuperRes is better than "simple" good sharpening, when starting with a good artifact-free upscaling algorithm. Still need to do some more investigation into the matter. SuperRes is a relatively "young" algorithm.

Simple good sharpening doesn't work with material that has noise and artifacts in it. SuperRes works GREAT to neutralize noise and artifacts while bringing out the real details. Even when using low-quality material with bilinear resizing, with SuperRes it still looks good. That can't be denied.

To the point that I really need to find a way to use SuperRes in AviSynth. madVR does a GREAT job but is limited by what my computer can handle in real-time. I developed a software that allows easy re-encoding of videos through AviSynth by applying EEDI3, NNEDI3, denoise, sharpening, SVP and other things as needed to greatly enhance the quality of the video; and then re-encode at higher quality. But... madVR has gone beyond what is available in AviSynth!

[MysteryX]

Hey I have an idea to simplify SuperRes. Instead of having all these settings and asking whether to apply once or every 2x scaling, why not simply ask how many passes he wants to run? If he runs 2 passes with a 3.5x upscaling factor, it will do 1 pass mid-way and 1 pass after. The user generally selects the number of passes based on what his graphic card can handle, but he really doesn't need to know when each pass actually gets applied. You can determine the best distribution of passes yourself.

[MysteryX]

One more thing. This is a bit off-topic but the tool I designed to re-encode videos is very useful to convert 288p or 360p into 720p, and then play it with madVR to fit your display. For really low-quality videos, it looked MUCH better than by simply running madVR. BUT, with SuperRes, it's almost even between my version and madVR's live version. My video is clearer with less noise, while the original file with madVR looks sharper because of SuperRes. I'd be happy to release that tool for any of you guys if interested (although the software itself was designed for something else, but has the encoder as part of it).

There are two issues I'm having problems with, however.

1. Playing .AVS (avisynth) script files in Windows Media Player ActiveX control, to preview the changes before encoding

2. Getting SuperRes to work in AviSynth (best bet would probably be editing AviShader which uses a shader in AviSynth)

You can contact me if you believe you can help with either of these issues.

[Hyllian]

Originally Posted by MysteryX

Personally I find the image unnatural. Too "cartoonish" for my taste. But then that probably works for Mario Bros

You may be right, in the sense that I, in fact, don't have a good knowledge of how to adjust a image of a video/movie (I'm completely newbie in this realm).

The algorithm is adjustable and I think you all video experts can extract much more from this than me in that task.

When you develop an edge detector algorithm, you tend to like edges a bit too much.