madVR - high quality video renderer (GPU assisted)

[Anima123]

As I remembered, Madshi has mentioned an idea of remove movie-like spots from the original picture, upscaling the cleaned picture then add the upscalied spots back to the upscaled cleaned picture. I wonder if this idea can be extended to something else, like noise or other effects?

[Arm3nian]

I did some testing using a normal quality source, I don't know how it was encoded as I just found it on youtube but it's obviously not bluray quality but it's not complete crap either. I think it's useful to have comparisons when the sources aren't on either extremes. It's pretty much just geometrical shapes so meh, madshi has probably done his own testing.

I compare regular NGU, nnedi3, and superxbr. I was going to test every combination, but realized 15min in that i'd be there until tomorrow. The pictures are mostly the highest levels you can select on each algo. nnedi3 and super-xbr are quadrupled - it makes quite a difference on nnedi3, but not that much on super-xbr. NGU levels for chroma make quite a difference. nnedi3 levels not really. NGU luma levels make even more of a difference, low to medium is quite the upgrade.

I tried pixart after, and it looks the worst for luma. However, it looks very similar to nnedi3 for chroma.

http://imgur.com/a/h2Imd

[StinDaWg]

madshi, I am still getting constant black screen lockups when existing/closing a video, that require a reboot of the computer. I have tried every combination I can think of. D3D9, D3D11, FSE and windowed, DontRenderAfterStop folder with and without.

Originally Posted by madshi

It even happens if you disable fullscreen exclusive mode!?!? That seems very weird. Which OS and GPU are you using? If it's NVidia or Intel, have you tried Overlay mode? Otherwise you could also try reducing the number of pre-presented frames ("present in advance"), or turning pre-present off completely.

Have you tried different GPU driver versions? This is most likely a GPU driver bug.

Windows 10. HD7770. I've used different drivers. I have had this problem on and off for years honestly. It effects both exclusive and windowed and all resolutions/refresh rate.

Currently I am using KeepLastFrameOnStop folder "hack" and it hasn't locked up for 2 days. I also checked "use a separate device for presentation" and "use a separate device for dxva processing".

I'll keep doing that and if it ends up locking up again I know I can move on and try something else (lowering present in advance as you suggested).

It really sucks but if there is no fix then I guess I'll just deal with it for better PQ.

[madshi]

Originally Posted by TheShadowRunner

Yes it happens with all DVDs I could try (10+).
To reproduce: just start any dvd playback, on the menu (if there's one) select the main movie or some extra. And then just seek during playback or jump to a random position : the player instantly becomes unresponsive (interestingly the bug doesn't happen when going from one chapter to another).

I've just tried with a ripped DVD sample a user sent me months ago, and I don't get any freeze here while seeking during playback.

Can you maybe create another sample for me with which I could reproduce the problem? Doesn't have to be a full 4-8GB DVD, you can probably cut the movie VOB down to a small size, so that maybe just the first 2 minutes or so are playable.

Originally Posted by Damien147

My TV{(Samsung Ku6000 (or else ku6300)}is 60hz and says in the manual that it supports film mode only for 1080i in HDMI.
In windows display adapter properties supported screen refresh rates are:24,25,29,30,50,59,60 hz.
I played with display modes in Madvr and it seems that 24p,25p,30p,50p is activated and displayed in 4096x2160(Madvr and TV report it).
If ''bypasses'' is the proper word do you guys think that what I am doing might cause problem/damage to the display?
I mean I haven't found support for that officially.

I don't understand your question. You're asking if what you're doing might cause problems, but you don't fully describe what you're doing exactly, and why you think that could cause problems.

So my last post (with the dozens of replies in it) to see some recommendations I made there for another Samsung 4K display user.

Originally Posted by Shiandow

Not that it's particularly important, but proper bilinear downscaling (i.e. a box filter) gives better results than bilinear interpolation for pretty much any scaling factor other than 2 (when they're equivalent).

Are you sure? I thought texture units would do proper bilinear downscaling for down scaling factors up to 2.0x? What are they doing if not proper bilinear downscaling?

Originally Posted by ryrynz

you're right the video I linked you before is a much better test of low res animated content.. only NGU very high does okay with the lines (everything else loses to NNEDI3 32 neurons) but suffers from the line issues I've mentioned previously.

In which way do NGU pixart low/medium lose to NNEDI3-32? More aliasing? Or other issues? In my tests NGU pixart medium seemed to have noticeably less aliasing than NNEDI3-16, at the very least. How do you like NGU pixart low/medium compared to NNEDI3-16? Both are faster than NNEDI3-16, so that's the fairest comparison, I would say.

Originally Posted by Bloodred

I've tried to find a few more scenes, though I think the one I first posted is one of the best examples, as in the image itself isn't of very poor quality and also showcases the issue. I've zipped a few images, just the unscaled versions:
NGU problematic scenes
Out of those images, the Slayers frame is by far the worst, but it's also very low quality, so it's to be expected.

One thing I've noticed is that the smudged/wavy look is most apparent when doing NGU image quadrupling, it's not much of an issue otherwise, so this particular usecase I ran into (upscaling 480p > 2160p) is what makes it apparent. Some of those shots aren't all too bad, but if you switch back and forth between NNEDI3 and NGU quadrupling it's easily visible. It's also most apparent around areas with a lot of black outlines, but it does happen in colored areas too.

I've used NGU vhigh across the board when looking for these scenes, with Jinc AR to upscale all the way to 2160p and deband set to Low.

Let me know if you could use even more images, I don't have a lot of low-res sources, but I'll try to look for more scenes where the issues are visible if you need more stuff to test with.

Thanks, I'll have a look at the images when I find some time!

Originally Posted by jkauff

I take it all back. I took several screenshots from the worst SD sources I could find, and NGU pixart High did a better job on every set (using NNEDI3 128 neurons and 256 neurons both).

I'm sold.

Originally Posted by MS-DOS

Both. SXBR+SR1 (mistake, sorry) against straight NGU. I didn't test NGU with SR, because apart from you not recommending it it's also very slow. Here's what I just got:

SXBR+SR = 4.92ms
NGU Med = 5.41ms

Nevermind it though, I probably shouldn't mention it in the first place as that's an issue with the AMD driver + RX 480 and not with NGU.

Ah, SR1 instead of SR3, and Polaris. That explains the speed stuff, of course.

Originally Posted by MS-DOS

I currently don't have a good video frame example to post, so here's a couple of images: 1 2. Double the first one and quadruple the second. With pixart, the pinned flag (or whatever it's called) on the first one looks aliased, while with the second one the whole image turns into an artificial mess with aliased edges.

Thanks, will have a look later.

Originally Posted by MS-DOS

With real life content I often need to scale bad quality sources with compression artifacts. Guess what NGU does? It enhances (sharpens) the quality of the artifacts while leaving the actual data visually almost untouched compared to SXBR+SR. Sometimes it produces aliasing, sometimes it just looks too artificial to me, and sometimes it looks better than SXBR+SR. Therefore it's not a universal upscaler that I would use for everything.

Well, standard NGU does not do good with compression artifacts, that's true. NGU pixart should handle such sources much better, though. And of course you can also apply SR with NGU pixart, but obviously that costs more GPU power.

Originally Posted by leeperry

^ Fully agreed, NGU doesn't quite get through noisy/low-res sources as NNEDI3 & sxbr do.....

You seem to have missed NGU pixart.

Originally Posted by Anima123

As I remembered, Madshi has mentioned an idea of remove movie-like spots from the original picture, upscaling the cleaned picture then add the upscalied spots back to the upscaled cleaned picture.

Huh? I'm not sure what you mean.

Originally Posted by Arm3nian

I did some testing using a normal quality source, I don't know how it was encoded as I just found it on youtube but it's obviously not bluray quality but it's not complete crap either. I think it's useful to have comparisons when the sources aren't on either extremes. It's pretty much just geometrical shapes so meh, madshi has probably done his own testing.

I compare regular NGU, nnedi3, and superxbr. I was going to test every combination, but realized 15min in that i'd be there until tomorrow. The pictures are mostly the highest levels you can select on each algo. nnedi3 and super-xbr are quadrupled - it makes quite a difference on nnedi3, but not that much on super-xbr. NGU levels for chroma make quite a difference. nnedi3 levels not really. NGU luma levels make even more of a difference, low to medium is quite the upgrade.

I tried pixart after, and it looks the worst for luma. However, it looks very similar to nnedi3 for chroma.

http://imgur.com/a/h2Imd

So what are your conclusions? I think real world content is more important than geometric figures, though.

[madshi]

Originally Posted by StinDaWg

Windows 10. HD7770. I've used different drivers. I have had this problem on and off for years honestly. It effects both exclusive and windowed and all resolutions/refresh rate.

Currently I am using KeepLastFrameOnStop folder "hack" and it hasn't locked up for 2 days. I also checked "use a separate device for presentation" and "use a separate device for dxva processing".

I'll keep doing that and if it ends up locking up again I know I can move on and try something else (lowering present in advance as you suggested).

It really sucks but if there is no fix then I guess I'll just deal with it for better PQ.

Well, what can I say? I don't have this issue, and most other users don't seem to have it, either. So I'm not completely sure why it occurs for you. I'd like to blame Windows 10, but if you had it for years, that can't be it, obviously.

[StinDaWg]

Originally Posted by madshi

Well, what can I say? I don't have this issue, and most other users don't seem to have it, either. So I'm not completely sure why it occurs for you. I'd like to blame Windows 10, but if you had it for years, that can't be it, obviously.

Do you hit the close button/backspace to end a video, or do you exit out of fullscreen first and then close the video program? What do you use? MPC-HC?

Just want to make sure we are on the same page here, but obviously what I am referring to happens on the former. I can exit out of fullscreen 1st and then close the player without issue, it's only when closing from fullscreen that it sometimes locks up.

Also I tried MPC-BE and same thing happened, so it eliminates the issue as being player specific.

[Shiandow]

Originally Posted by madshi

Are you sure? I thought texture units would do proper bilinear downscaling for down scaling factors up to 2.0x? What are they doing if not proper bilinear downscaling?

The GPU just interpolates, it has no clue if it's downscaling or not, this means it uses at most 2 pixels in each direction to calculate the interpolated value (so 4 in total). Now even the narrowest possible kernel for downscaling, a box filter, can use as many as 9 input pixels to calculate a single output pixel. This is because any pixel larger than the original pixel can overlap up to 9 of the original pixels. Now it's debatable what the proper downscaling analogue of bilinear interpolation is, but it's not just bilinear interpolation, not even for scaling factors <2.

[madshi]

Originally Posted by StinDaWg

Do you hit the close button/backspace to end a video, or do you exit out of fullscreen first and then close the video program? What do you use? MPC-HC?

Just want to make sure we are on the same page here, but obviously what I am referring to happens on the former. I can exit out of fullscreen 1st and then close the player without issue, it's only when closing from fullscreen that it sometimes locks up.

Also I tried MPC-BE and same thing happened, so it eliminates the issue as being player specific.

I'm using a custom keyboard shortcut to close MPC-HC while it's in FSE mode. Works fine here.

MPC-HC and MPC-BE are closely related. You'd need to test a totally different player to make sure it's really not media player specific.

Originally Posted by Shiandow

The GPU just interpolates, it has no clue if it's downscaling or not, this means it uses at most 2 pixels in each direction to calculate the interpolated value (so 4 in total). Now even the narrowest possible kernel for downscaling, a box filter, can use as many as 9 input pixels to calculate a single output pixel. This is because any pixel larger than the original pixel can overlap up to 9 of the original pixels. Now it's debatable what the proper downscaling analogue of bilinear interpolation is, but it's not just bilinear interpolation, not even for scaling factors <2.

You're right, of course, thanks.

[leeperry]

Originally Posted by madshi

You seem to have missed NGU pixart.

The thing is that I like the slight EE sxbr gives, especially in combination with SR as they look amazing together on fast-motion 720p movies IME and NGU is slower than usual on AMD 480's apparently? It also isn't SR-friendly due to excessive bloating and we need horse-power for 2160p upscaling you know.

Anyway, please don't leave us NGU as the only option because as much as we don't miss NEDI I can certainly see some NNEDI3/sxbr lovers and again options are good, we got half a dozen linear scalers to pick from at every step so what's the big deal with three doublers

[StinDaWg]

Originally Posted by madshi

I'm using a custom keyboard shortcut to close MPC-HC while it's in FSE mode. Works fine here.

What's your shortcut? And is it set to "close" or "exit"?

I could try it and see if it makes any difference.

[madshi]

Originally Posted by leeperry

I can certainly see some NNEDI3/sxbr lovers

You're missing the mark. NGU pixart is aimed at replacing NNEDI3 1:1. The goal is to achieve the same look and quality at faster speed. Whether or not NGU pixart is there just yet is a matter of discussion. But once it is, why would anyone use an algorithm that doesn't look better, is slower and requires OpenCL with its unstable drivers?

This is about NNEDI3, though. super-xbr clearly has a different look to it and NGU pixart is not meant to reproduce the super-xbr look. However, most users probably consider NNEDI3 to have better image quality than super-xbr. So once something is available at equal or higher speed than super-xbr with better image quality, arguments for keeping super-xbr might be hard to find. But again, maybe we're not there just yet.

Originally Posted by StinDaWg

What's your shortcut? And is it set to "close" or "exit"?

I could try it and see if it makes any difference.

I'm using Ctrl+C, but not to close but to exit. By default Ctrl+C closes. I've changed it to exit instead.

[Damien147]

Originally Posted by madshi

I

I don't understand your question. You're asking if what you're doing might cause problems, but you don't fully describe what you're doing exactly, and why you think that could cause problems.

So my last post (with the dozens of replies in it) to see some recommendations I made there for another Samsung 4K display user.

I followed these recommendations already.
From TV what I get officially is film mode ONLY for 1080i in HDMI,there's no mention for general 24p support or anything.Everything's supposed to play at 60hz and film mode is greyed out.
By playing with Madvrs display modes ''it bypasses'' what I explained and I have 24fps at 24hz, 25fps at 25hz and beyond.... in 4096x2160 resolution with everything I put.
There's no official support for that but it seems to work.
Here what Madvrs OSD reports
http://imgur.com/a/gpakS

and here what TV reports
http://imgur.com/a/AzCMC

It's not supposed to happen officially but it happens and I'd like to keep it like that.The thing is,am I gonna have a problem/damage the display in the long run for some reason?

[madshi]

Originally Posted by Damien147

I followed these recommendations already.

No, I don't think you did. Or is 4096x2160 the physical resolution of your display? I doubt it. Most 4K displays are 3840x2160 and not 4096x2160. So you probably did not follow my recommendation to use the physical resolution of your display. Or am I wrong?

Originally Posted by Damien147

The thing is,am I gonna have a problem/damage the display in the long run for some reason?

I don't think it's easy or even possible to damage digital displays by sending unusual resolutions/refresh rates. If your TV doesn't like the signal you're sending, it will not sync. If it syncs, there should be no problem.

[huhn]

"all" UHD TV support an input of 4096x2160 and it is not the native resolution only projector have this resolution.

he send 4096x2160 the TV is downscaling it to 3840x2160 by using the fit to screen option without this option it is cropping the image (which is the correct way to handle it unlike fit to screen.)

of cause 4096x2160 is supported but it is suboptimal to say it very friendly.

[Damien147]

Originally Posted by madshi

No, I don't think you did. Or is 4096x2160 the physical resolution of your display? I doubt it. Most 4K displays are 3840x2160 and not 4096x2160. So you probably did not follow my recommendation to use the physical resolution of your display. Or am I wrong?

It is 4096x2160.

Originally Posted by madshi

I don't think it's easy or even possible to damage digital displays by sending unusual resolutions/refresh rates. If your TV doesn't like the signal you're sending, it will not sync. If it syncs, there should be no problem.

Thank you and your great software then.

Originally Posted by huhn

"all" UHD TV support an input of 4096x2160 and it is not the native resolution only projector have this resolution.

he send 4096x2160 the TV is downscaling it to 3840x2160 by using the fit to screen option without this option it is cropping the image (which is the correct way to handle it unlike fit to screen.)

of cause 4096x2160 is supported but it is suboptimal to say it very friendly.

If I put 2160p24 for example in display modes with an 3840x2160 movie what you say happens,it downscales in 3840x2160 and TV reports it.If I put 4096x2160p24 for the same movie it stays 4096x2160.
Fit to screen is 1:1 pixel mapping.

[madshi]

Originally Posted by Damien147

It is 4096x2160.

Oh, that surprises me. Are you sure? Which TV is this exactly? I was only aware of projectors (Sony, JVC) with 4096 wide panels.

[StinDaWg]

Originally Posted by madshi

I'm using Ctrl+C, but not to close but to exit. By default Ctrl+C closes. I've changed it to exit instead.

OK, I set exit to F4, which is what my media center remote is mapped to the close button.

Before "close' was mapped to "close", now it's mapped to "exit". We'll see if that makes any difference as they seem to both do the same thing.

[huhn]

Originally Posted by Damien147

If I put 2160p24 for example in display modes with an 3840x2160 movie what you say happens,it downscales in 3840x2160 and TV reports it.If I put 4096x2160p24 for the same movie it stays 4096x2160.
Fit to screen is 1:1 pixel mapping.

fit to screen is the opposed of 1:1 pixel mapping.

select the resolution 2048x1536 and you may understand it.

[Damien147]

Originally Posted by madshi

Oh, that surprises me. Are you sure? Which TV is this exactly? I was only aware of projectors (Sony, JVC) with 4096 wide panels.

Samsung KU6000 or else KU6300.

[strumf666]

http://www.samsung.com/uk/tv/KU6000/

"Resolution
3,840 × 2,160"