madVR - high quality video renderer (GPU assisted)

[Asmodian]

Any setting except always is reasonable in my opinion, I usually use 1.5. Below that and I don't think NNEDI3 is worth it.

2.0 is also a good option, with 2.0 you only use NNEDI3 if you will not have to downscale after it.

[fairchild]

Originally Posted by Asmodian

2.0 is also a good option, with 2.0 you only use NNEDI3 if you will not have to downscale after it.

I prefer using 2.0 as well as I don't think the benefit of having to upscale past your display resolution only to then also have to downscale will be all that beneficial. Though some swear by using NNEDI3 doubling with 1280x720 videos to 1920x1080 display. I think it's just placebo for these people with their super duper binocular eyes.

[burfadel]

I have a problem with the on-screen controls with my connected Samsung H6400 TV. When playing MPC-HC, the on screen controls do not show in full screen! However, this is not always the case.

The problem modes are:
1920x1080 (both 50 and 60 Hz). This is the mode I actually want to use, the desktop fits perfectly on the screen. You should also expect it to work in this mode since it's default!
Any custom made through AMD Catalyst Control Centre (including 1776x1000)

It is not a scaling issue (that is, the controls are not off the edge of the screen). They are simply non-existent. If you click where the controls should pop up, it just pauses. If I underscan the image in Catalyst Control Centre, the desktop of course has black borders around the edge of the screen, the controls simply are not there.

The modes where the on-screen controls work:
1280x720
1776x1000 optimised (this is mode listed in the HDTV support section). So, for some reason 1776x1000 works, but not a custom 1776x1000, nor 1920x1080.

I could of course just use 1776x1000, HOWEVER the picture does not fill the whole screen. This is kind of obvious with it being 1776x1000, it basically shows 1776x1000 inside a 1920x1080 window, essentially. Basically, I have black borders around the image. As this is a custom mode, I cannot use the scaling options! The amount of border is quite large considering it is a 55 inch TV.

What I have tried:
Changing the output colour depth in the AMD Control Centre, which is under Properies (Digital Flat-Panel) between 8bpc, 10bpc, and 12bpc (yes, 12bpc is an option on this screen also).

Playing around with the LAVF settings, output colour spaces etc. No luck.

Several other things!

Has anyone else experience this? Can it be fixed? It is super annoying!

[huhn]

is this happening with EVR too?

this sounds like a problem with mpc hc not madvr

[burfadel]

Okay, I found the issue. It is a bug in MPC-HC! It's working properly now, in fact I can't make it not work properly!

How to 'fix'

Make sure the monitor (or TV etc) is in it's native mode, say 1920x1080. Also make sure MadVR etc is selected, then:

• go to view --> options
• click on 'Fullscreen' under 'Playback' on the left hand side
• select the correct monitor from the 'Fullscreen monitor' dropdown list
• click on 'Use autochange fullscreen monitor mode
• click add, which should add an option to the list at the bottom with the correct display mode and refresh rate
• make sure the 'from' and 'to' are set correctly (for this purpose type in say, from 1 to 70, it doesn't matter, as long as it covers what you will play)
• move it up to the top and deselect every other option
• deselect the 'Apply default monitor mode on fullscreen exit' and 'Restore resolution on program exit' at the bottom
• click ok
• open a new video and play in fullscreen mode
• check out the working controls
• exit fullscreen mode
• go back to options and deselect the autochange option, it seems to be only needed once
• close MPC-HC
• reopen and test with your video

If it wasn't working before, it should now. Seems there is a settings bug in MPC-HC somewhere. It works after a reboot etc as well as changing display modes, which is good!

[StinDaWg]

Originally Posted by fairchild

I prefer using 2.0 as well as I don't think the benefit of having to upscale past your display resolution only to then also have to downscale will be all that beneficial. Though some swear by using NNEDI3 doubling with 1280x720 videos to 1920x1080 display. I think it's just placebo for these people with their super duper binocular eyes.

Not placebo. There is way less artifacts and a sharper image using NNEDI3 on 720p->1080p.

[leeperry]

Originally Posted by sheppaul

Probably you are using Direct3D 9 skin mode?

Oh you nail it, I didn't see this one coming!

If madshi could do anything at all, this would be much appreciated. For now, I'll unassign the pause button in my custom skin....the auto-resume feature will still get the job done.

[tjcinnamon]

Originally Posted by StinDaWg

Not placebo. There is way less artifacts and a sharper image using NNEDI3 on 720p->1080p.

What is the setting you use?

always
1.2
1.5
2.0?

[QBhd]

I do 720p==>1024x768 and have it set to "always" and I can say for a certainty that there is no placebo effect going on... I too was skeptical, until one day when I decided to tick that radial. And the first thing that popped in my head was "Wow!"... So if I had a 1080 display, I would always be using NNEDI3 for image upscaling.

QB

[leeperry]

Originally Posted by QBhd

I would always be using NNEDI3 for image upscaling.

I think StinDaWg & Frechdachs nailed it coz what I don't like with NNEDI is the aliasing they're reporting with AR+LL, will need to reevaluate.

[QBhd]

Well if you look at the madVR settings, AR does increase the aliasing bar... as for LL, I am fairly sure it is only supposed to be used with Catmull-Rom

[Stereodude]

Originally Posted by burfadel

Has anyone else experience this? Can it be fixed? It is super annoying!

It sounds like you need to change the TV to the 1:1 pixel setting that disables overscan.

[trip_let]

Originally Posted by tjcinnamon

What is the setting you use?

always
1.2
1.5
2.0?

To make it active on 720p -> 1080p it can't be 2.0.

Originally Posted by StinDaWg

Not placebo. There is way less artifacts and a sharper image using NNEDI3 on 720p->1080p.

Depends on content a lot and downscaling used, etc.

Here's a quick and dirty look though. Source was posted a while back.

720p -> 1080p
Catmull-Rom
Lanczos 3 AR
NNEDI3 16 neurons doubling, Mitchell-Netravali LL downscale

Okay, not quite comparable or fair maybe and not the best choices, but that's what I had saved prior in a quick test for myself. The text at the bottom looks appreciably better with the NNEDI3. I guess many would use a sharper downscaling algorithm though.

For most images most of the algorithms look pretty similar on just 1.5x though, unless pixels are large. I don't know if I'd be taken by hyperbole on describing these things.

[cyberbeing]

NVIDIA users should beware that the 340.43 Beta seems to have a bug which causes Custom Resolutions created by previous drivers to be ignored. They must have accidentally or otherwise changed the binary format of the CustomDisplay registry key slightly, making this new driver think the old key is corrupt. You can create them from scratch again, but rather annoying if you had a lot of special custom timings. YMMV.

[StinDaWg]

Originally Posted by leeperry

I think StinDaWg & Frechdachs nailed it coz what I don't like with NNEDI is the aliasing they're reporting with AR+LL, will need to reevaluate.

I'm still waiting to hear what madshi has to say about it, just going to use without AR for now.

Originally Posted by QBhd

Well if you look at the madVR settings, AR does increase the aliasing bar...

Ya, but it's a HUGE difference, and only happens with downscaling. No AR text looks 99% perfect, AR on text looks really jagged (worse than just Jinc3AR upscaling).

The first image in each is Lanczos4, the 2nd is Lanczos4 AR. Look at letters like M,N,V,A.


http://screenshotcomparison.com/comp...9590/picture:0

[cyberbeing]

StinDaWg, do you see the same when using Spline 3-tap AR upscaling/downscaling, NNEDI3 256 chroma upscaling, NNEDI3 256 luma doubling?

Testing with 256 neurons is to rule out AR enhancing an NNEDI3 artifact. NNEDI3 has a tendency to produce blocky artifacts on high contrast text edges in particular. Only 256 neurons is completely immune.

Testing with Spline 3-tap is again to rule out potential anomalies. madVR's AR algorithm used with all resizers except Jinc was originally tuned to be near optimal (>95% eliminated ringing, without loss of detail loss or sharpness) only with Spline 3-tap. AR has always produced residual ringing artifacts with Lanczos3/4/8, Bicubic75/100, Spline4, and Jinc3/4 which exponentially become more visible with higher taps. The low ringing and soft algorithms on the other hand experience a slight loss of detail and sharpness with AR.


I'd also be curious if you can you reproduce this with madVR v0.87.8, which is prior to madVR correcting NNEDI3's 0.5 pixel shift when performing additional scaling. When pixel peeping I noticed that the text in your AR screenshots appears shifted slightly left & up compared to the non-AR screenshots, which is a bit strange.

[seiyafan]

So is it still recommended to use Catmull-Rom AR LL for downscaling with NNEDI3 or should it be Spine 3?

[6233638]

FYI I have responded to the discussion of anti-ringing/linear light scaling/NNEDI3, but a moderator has split it off into its own topic here: http://forum.doom9.org/showthread.php?p=1684179

[StinDaWg]

Originally Posted by cyberbeing

StinDaWg, do you see the same when using Spline 3-tap AR upscaling/downscaling, NNEDI3 256 chroma upscaling, NNEDI3 256 luma doubling?

Testing with 256 neurons is to rule out AR enhancing an NNEDI3 artifact. NNEDI3 has a tendency to produce blocky artifacts on high contrast text edges in particular. Only 256 neurons is completely immune.

Testing with Spline 3-tap is again to rule out potential anomalies. madVR's AR algorithm used with all resizers except Jinc was originally tuned to be near optimal (>95% eliminated ringing, without loss of detail loss or sharpness) only with Spline 3-tap. AR has always produced residual ringing artifacts with Lanczos3/4/8, Bicubic75/100, Spline4, and Jinc3/4 which exponentially become more visible with higher taps. The low ringing and soft algorithms on the other hand experience a slight loss of detail and sharpness with AR.


I'd also be curious if you can you reproduce this with madVR v0.87.8, which is prior to madVR correcting NNEDI3's 0.5 pixel shift when performing additional scaling. When pixel peeping I noticed that the text in your AR screenshots appears shifted slightly left & up compared to the non-AR screenshots, which is a bit strange.

I remember first seeing this issue when NNEDI3 was enabled, so I don't think madVR versions has anything to do with it. I tried NNEDI3 256 neurons/Spline 3AR and it isn't any better, still aliased text. The only thing that removes aliasing on all downscaling algorithms is to not use AR.

Originally Posted by seiyafan

So is it still recommended to use Catmull-Rom AR LL for downscaling with NNEDI3 or should it be Spine 3?

I've always wondered why it was recommended here to use Catmull-Rom for downscaling. In various encoding forums where people are obsessed with quality, everyone uses Spline 36 (I guess this is Spline 3 in madVR). Spline is almost as sharp as Lanczos with less ringing/artifacts. Catmull-Rom is too soft IMO. Spline is somewhere in the middle.

If the aliasing is inherent to the algorithm and can't be fixed, I don't think AR should be recommended to use anymore for downscaling (it's still fine for upscaling). One of the greatest benefits from NNEDI3 is near perfect diagonal lines, using AR negates that. I don't see any increased ringing from not using it anyways, at least for 720p->1080p.

Edit: Just took a look at the pictures posted in that linked post and the issue definitely seems to have something to do with the way NNEDI3 doubles. Don't know why it was broken off into another thread either as it's relevant to this discussion we're having.

[Guest]

The moderating strategy was discussed with and approved by madshi, who is the originator of this thread. Please trust us in our attempts to keep things organized and accessible. It is not just a question of relevance. We have 26000+ posts in this thread! It is a nightmare to find things and very hard for madshi. Please continue the topic of linear light scaling/NNEDI3 in the linked thread. Thank you for your understanding.

Just to remind about the policy: This thread is for support of madVR (bug reports, questions about usage, etc.) and occasional development things when madshi asks for them (help with testing, etc). Everything else including theoretical discussions should go into a separate thread. You can preface the thread with "madVR" so it is recognizable as relevant to madVR.