madVR - high quality video renderer (GPU assisted) [Archive] - Page 1276

flossy_cake

16th January 2023, 02:32

sorry it is really trivial if you ignore repeat flags in a soft telecine file you just end up with 23p you will not run async it's not that complicate.

For a 'clean' source, yes. But so many NTSC DVDs are 'dirty' and has all these edits done in the 29.97i domain, like the intro sequences, overlayed credits and orphan fields on scene changes. Sometimes the fps changes to 29.97 just for those sequences, other times it stays at 23.976 but those frames are combed - they are flagged as interlaced in MPC video renderer debug screen with "i" next to the frame counter. My sony DVD player will deinterlace only those frames, and leave the rest as progressive. In Avisynth the frame property _FieldBased changes from "progressive" to "tff/bff" just for those frames (this is only available with LwLibAvVideo source filter - DirectShowSource and FFMS2 don't set the _FieldBased property at all).

I tried to tolerate the combed frames for those sequences but they are bothersome to me. NVidia/AMD deint is "video mode" and lose half res on movement, so I didn't want that either. For now I'm using Avisynth TFM & TDecimate to soft telecine -> hard telecine -> IVTC & deint those interlaced frames. If I leave it as soft telecine and just deint only the _FieldBased=progressive frames, then I end up with repeat frame on orphaned field scene splices which is noticeable to me (this happens because the top field of that frame is used for deinterlacing, and that top field contains same image as previous IVTC'd frame).

The last of those 3 videos is tricky as it uses 3:3:2:3:2 cadence which is actually 22.976p after IVTC. For that I tried using Avisynth AssumeFPS(24000, 1001, true) to speed it up to 23.976 kind of like PAL speedup and then pitch correct the audio with TimeStretch(pitch=((23000.0/1001.0)/(24000.0/1001.0))*100) which worked for most of the episode, but then there would be certain sequences that change to some other weird cadence like 2:2:3:2 which become stuttery.

cjfournier

17th January 2023, 02:00

Having a weird issue I believe others have seen, but haven't yet found any comments pointing to solution: Currently using an AMD R9 390 w/ driver 22.6.1 (have tried several drivers, issue is the same across all) on an Optoma projector with active shutter glasses; When playing 3D in full screen (windowed mode is fine), the video starts fine but begins to "drift".

What's weird is that video looks normal through the right lens, but closing the right eye has the left side cycling through a set of static images. However, when something is brought to the foreground, like a properties window or progress bar, the video catches up and all looks normal. I believe I've toggled through just about every madVR option, but nothing seems to influence the behaviour. The same was described by another user over on the Jriver forum, but there was no mention of what card was used (guessing AMD, like me.)

Band-aid is to keep a properties window open, but shoved off to the lower corner.

Is this a known condition, and is there a proper fix?

Sunspark

17th January 2023, 04:46

I have never used this hardware, but I wonder what happens when you try other renderers like Microsoft Movies & TV app, VLC, etc.

thighhighs

17th January 2023, 05:53

Hello. Is the anti-bloating filter useful for SSIM1D?
I'm talking about "real" downscaling 2160 blu-ray -> 1080. I see opposite results when downsizing 2160 blu-ray and hi-res images (~2160 or more). For images 1D100 + AB25% works well: better sharpness compared to Bicubic150, and better overall look compared to SSIM1D no-AB. But for 2160 video Bicubic150 is starting to look better: it more sharp compared to SSIM1D100+25%AB. I don't understand what's going on anymore :confused:

ryrynz

17th January 2023, 07:40

Hello. Is the anti-bloating filter useful for SSIM1D?

Nothing to be confused about, you decide this for yourself based on what you see with your own eyes, don't get hung up on it.
Bicubic is a very popular downscaling algorithm and you can see why.

thighhighs

17th January 2023, 09:18

ryrynz
I would like to know what other users think about 1D. I think 25% AB sometimes helps for video 2160 -> 1080, but comes at a high price.
By the way, as far as I remember your posts, you really like SSIM2D. What combination of SSIM2D/AB are you using? This algorithm I also used and would like to know more opinions. Especially from those who actually use it.

flossy_cake

17th January 2023, 14:38

Take some comparison screenshots on the same frame then zoom in like 400% nearest neighbour to see what's going on with the pixels. I find the debloat has quite a softening effect when used with the sharpeners.

In an earlier post I was raving about how much I liked AdaptiveSharpen and LumaSharpen but after looking at lots of SD content over the past few months I'm now preferring the sharpen-crispen-thin-enhance combo (with hard upper limit of 1.0 for the last 2 otherwise they are too artefacty for my liking).

kopija

17th January 2023, 17:19

Greetings,
I intended to buy a Freesync monitor thinking that would remove need for smoothmotion/resolution autochange/custom resolution workarounds for eliminating judder.
But now I feel like I stepped into a minefield after doing some research.
First, Madshi saying MadVR does not support Freesync, then numerous posts saying it works.
It also supposedly works with major video players, (including Windows Movie Player!)
I really like Madvr black bar removal feature.
Is there any way to make Madvr work with Freesync?
I originally intended to buy 1080p 75Hz Freesync monitor, thinking that it would auto-switch to 72/60/50Hz and eliminate need for abovementioned workarounds.
Can Freesync do that? Are there any known downsides in such a scenario?
I know that 144Hz monitor running fixed 120Hz would be ideal for 24/30p, but it is out of my price range.
Thanks

thighhighs

17th January 2023, 18:08

Take some comparison screenshots on the same frame then zoom in like 400% nearest neighbour to see what's going on with the pixels.I'm not sure, I would like to know the opinion of others. For tests, I prefer to look from a normal distance. This is the only way I noticed how dramatically SSIM2D changes the overall look. Also, this is the only way I've noticed that 1D100 makes film grain more visible to my eyes. Even when the Bicubic150 and 2D are sharper, they don't have this problem, only the SSIM1D100 + no-AB has this problem.
I find the debloat has quite a softening effect when used with the sharpeners.The same goes for downscaling. AB can help or just reduce the sharpness...
A small example. SSIM2D100 + AB100 looks good in most cases and is much better than Bicubic150 or 1D. But in some cases it looks blurry.
https://imgsli.com/MTQ3OTU5
In this scene from The Hobbit, Bicubic150 is preferable to my eyes. That's why I don't like the combination of 2D with AB above 75%.
What do others think? What combination of SSIM1D/2D and AB are you using? Is any AB required for SSIM1D? What do you think?

Sunspark

17th January 2023, 22:05

kopija: You should see if your monitor can do 23.976/24/25 refresh rates. Mine is able to on the HDMI input. If it is able to, then you never need to use smooth motion, etc.

ryrynz

18th January 2023, 01:28

What do others think? What combination of SSIM1D/2D and AB are you using? Is any AB required for SSIM1D? What do you think?

I don't downscale any more as I have a 4K set, but 2D was quite superior to 1D. After my initial review of AB I decided I didn't find it useful, better imo to just reduce the sharpness of the display, but it all depends.

flossy_cake

18th January 2023, 02:08

1D100 makes film grain more visible to my eyes.

The "enhance detail" sharpener kind of enhances that too, along with texture details like wood grain will look more detailed. But the downside is it also enhances compression artefacts, so I don't use a value more than about 0.5 for compressy SD and about 1.0 for a very high quality DVD remux.

I'm quite happy with Lanczos for scaling and then using sharpeners if the source is soft. To me that makes more sense to do the sharpening separately as it gives me more control over the type of sharpening for the source (using profiles for convenience). Whereas I'm not sure how 1D100 might react with certain source types so I'd have to do a lot of trial and error.

I'm not really seeing much meaningful difference in your Hobbit comparison. Small differences, but nothing that would cause me to pick one over the other. Maybe I need to look under better lighting.

Greetings,
I intended to buy a Freesync monitor thinking that would remove need for smoothmotion/resolution autochange/custom resolution workarounds for eliminating judder.
But now I feel like I stepped into a minefield after doing some research.
First, Madshi saying MadVR does not support Freesync, then numerous posts saying it works.
It also supposedly works with major video players, (including Windows Movie Player!)
I really like Madvr black bar removal feature.
Is there any way to make Madvr work with Freesync?
I originally intended to buy 1080p 75Hz Freesync monitor, thinking that it would auto-switch to 72/60/50Hz and eliminate need for abovementioned workarounds.
Can Freesync do that? Are there any known downsides in such a scenario?
I know that 144Hz monitor running fixed 120Hz would be ideal for 24/30p, but it is out of my price range.
Thanks

I discussed that with Madshi some years ago and was begging him to officially support it when I first got my gsync monitor. As far as I'm aware there were no lines of code in MadVR written to facilitate gsync/freesync, but I was able to get it working in full screen mode on my Acer XB271HU (monitor's own gsync fps overlay in top right corner would match the content fps). But it had nothing to do with MadVR settings, it was entirely up to the DirectX version, being in the right full screen mode, and selecting the right NVidia Control Panel settings to make it work. I wouldn't rely on it working correctly or anything, but it probably would work in a certain configuration as long as the driver knows you are in full screen graphics mode or whatever the driver requires for freesync/gsync to become active.

huhn

18th January 2023, 06:17

I tried to tolerate the combed frames for those sequences but they are bothersome to me. NVidia/AMD deint is "video mode" and lose half res on movement, so I didn't want that either

but they are full res when working. that's why they are so good mixed content not perfect at all but good.Greetings,
I intended to buy a Freesync monitor thinking that would remove need for smoothmotion/resolution autochange/custom resolution workarounds for eliminating judder.
But now I feel like I stepped into a minefield after doing some research.
First, Madshi saying MadVR does not support Freesync, then numerous posts saying it works.
It also supposedly works with major video players, (including Windows Movie Player!)
I really like Madvr black bar removal feature.
Is there any way to make Madvr work with Freesync?
I originally intended to buy 1080p 75Hz Freesync monitor, thinking that it would auto-switch to 72/60/50Hz and eliminate need for abovementioned workarounds.
Can Freesync do that? Are there any known downsides in such a scenario?
I know that 144Hz monitor running fixed 120Hz would be ideal for 24/30p, but it is out of my price range.
Thanks
video is different from games.
you don't want to show the frame as fast as possible you want to show it for a very exact time .

is it possibile to make VRR work for video yes absolutely but that's unlikely to happen with madVR taking current development.
even HDMI has VFR for video now.

BTW.75 hz freesync devices are a pretty much fake usually having a range of 48-75 hz to make freesync work making it useless. low end devices are not rarely utterly broken too.
even freesync on LG OLED TVS is objectively broken and they are not cheap.

Sunspark

18th January 2023, 07:58

kopija

18th January 2023, 08:01

BTW.75 hz freesync devices are a pretty much fake usually having a range of 48-75 hz to make freesync work making it useless. low end devices are not rarely utterly broken too.
even freesync on LG OLED TVS is objectively broken and they are not cheap.

So the best way to go about judder elimination issue would be to buy monitor with 48-75Hz range and use "switch to matching display mode" in madVR for 24/25/30fps?

Also, there is this new thing called VESA MediaSync (https://www.anandtech.com/show/17367/vesa-launches-variable-refresh-display-perf-standards-adaptivesync-mediasync/2). It claims its main purpose is elimination of 3:2 pulldown.
Gimmick or gamechanger?

nevcairiel

18th January 2023, 10:00

MediaSync is not a new feature, just a standard set of tests that displays can be certified against, which is far more lenient then the gaming tests.

That does not change the situation that video presentation APIs on Windows are not designed to work with what we need for video - exact presentation at a pre-determined time.

huhn

18th January 2023, 11:13

sleep time the render call and pray that it works :-)

So the best way to go about judder elimination issue would be to buy monitor with 48-75Hz range and use "switch to matching display mode" in madVR for 24/25/30fps?

Also, there is this new thing called VESA MediaSync (https://www.anandtech.com/show/17367/vesa-launches-variable-refresh-display-perf-standards-adaptivesync-mediasync/2). It claims its main purpose is elimination of 3:2 pulldown.
Gimmick or gamechanger?

LG OLED are freesync premium and gsync certified.
so you can get the claim on most devices.

flossy_cake

18th January 2023, 11:19

flossy, I have an mkv that is a bt.601/pal/50 fps from a dvd. It's poor quality and the original source is from decades ago.

Resolution is 720x576. I decided to set the display resolution to 720x576 and just let the monitor upscale the image to the screen dimensions which is the same process I follow for 1080p and it works well as madvr is only handling the chroma in that case.

However, what I am seeing here is that there still needs to be some image scaling in the x dimension because of the whole square vs rectangular pixel aspect ratio. If I select square, it will turn off madvr scaling but it will squish in the sides a little. If I leave it on default/4:3 it will look more normal, but then x dimension scaling comes into play..

What image upscaling algo are you using in the situation where you are only scaling in X and only for the pixel not a different resolution? Catmull-rom looks ok, and looking at the formula the b value is 0, so it's probably a good choice.. wonder if there's a better choice?

PS. Have you tried your bt.601/PAL DVDs in VLC? I'm looking at the mkv in VLC and the colours and gamma are different from how madvr is doing it--richer and not as washed out. Interesting.. twiddling about with the VLC renderer choices, selecting opengl or d3d9 makes it washed out, but selecting d3d11 or auto makes it darker and richer like how they should look.

Can you send the file so I can check it out? I need to see what MadVR chooses for aspect ratio, matrix, primaries & limited range on its debug screen. It will also say whether it's guessing or getting that info from upstream decoder.

If it's a straight MPEG2 remux of a 4:3 DVD then debug screen should say 4:3, BT.601, EBU/PAL, limited range. If it was transcoded by someone else, then whatever software they used could have fudged something, depends how they encoded it.

I reckon it might be slightly safer to output square pixels over HDMI as I know the math MadVR uses is good on that. SD signals over HDMI have been dodgy for me in the past. If the show is 4:3 then you don't want to be pillarboxing that in the middle of 576p as the active pixel area will be only 540 wide which is lower than the source. The chroma shouldn't have issues as long as the final output to TV is 4:4:4 and therefore avoiding multiple 4:2:0 -> 4:4:4 conversions.

I'm satisfied with Lanczos scaling for everything, except chroma I use bicubic old because it happens to smooth out chroma artefacts in any sources which have such issues. But in a double blind test on a normal quality source I can't really tell any diff. If I pause on a scene & Ctrl+Alt+Shift+F12 to Cb/Cr channel and flick between chroma algos on SD source then yeah I can really see a lot of diff.

Sunspark

18th January 2023, 22:19

The reason I have done the png screenshots like this is because for whatever reason, if I take a screenshot with vlc it has the gamma/colour different than when it's in the player so they couldn't be used. This is just to show the difference in the players during playback side-by-side.

Usually I use overlay mode, but switched to windowed so I could capture the image.

Average rendering time can be ignored, it's only high because it's paused.

https://dl.dropboxusercontent.com/s/binlyveqfd9grbh/1-intro.png

https://dl.dropboxusercontent.com/s/84z8rhwbemydp6m/2-light.png

https://dl.dropboxusercontent.com/s/fkna5v4bz3kijyc/3-dark.png

flossy_cake

19th January 2023, 03:50

Yes I see what you mean - VLC is rendering with a higher gamma it seems.

Check what settings you've got under MadVR's "devices > monitor > color & gamma"

I'm not familiar with VLC but maybe it assumes by default that monitor gamma is 2.2 and SDR content is mastered at 2.4, therefore applying a +0.2 gamma shift by default - are there any settings in VLC which relate to this?

If I take your madvr screenshot and darken gamma by 0.2 in photoshop it appears quite close to the VLC one (animated png):

https://i2.lensdump.com/i/RxO2yr.png

So basically if your monitor's gamma is 2.4 then VLC is too dark (it will be giving you about 2.6).

If your monitor's gamma is 2.2 then VLC is roughly correct, but only on the assumption that most content - especially of that era of television - would have been mastered on 2.4 monitor.

VLC's behaviour can be reproduced with madvr's color & gamma settings if you desire. Just keep in mind that room lighting affects perceived gamma due to "simultaneous contrast" effect - in a daylit room 2.2 can look similar to 2.4 in a dim surround.

x7007

19th January 2023, 04:52

remind me again how do we get Dolby Vision? which player and codec do I need?

Sunspark

19th January 2023, 05:18

Yes I see what you mean - VLC is rendering with a higher gamma it seems.

More than that.. the colours themselves are different.. if you look at the red and blue in the first row, I guess similar to the same issue you had with Family Guy, except it doesn't have the blue cast. Unless the different shade of red, etc. is due only to gamma? Might be it, more below.

Check what settings you've got under MadVR's "devices > monitor > color & gamma"

Everything is at 0, and enable gamma processing is unchecked.

I'm not familiar with VLC but maybe it assumes by default that monitor gamma is 2.2 and SDR content is mastered at 2.4, therefore applying a +0.2 gamma shift by default - are there any settings in VLC which relate to this?

No idea, but what I have observed is that there is a choice of different renderers in VLC.. d3d11, d3d9, opengl, etc. d3d9 & opengl renderers do the washed-out effect but the d3d11 one doesn't. Not sure why there is a difference in presentation with d3d11.

So basically if your monitor's gamma is 2.4 then VLC is too dark (it will be giving you about 2.6).

It's actually set to "Mac Gamma" in the monitor settings instead of PC because I found PC to be too dark. Probably would be better if it was calibrated, but Mac Gamma 1.8 (maybe up to 2.0?, not sure how accurate it is) looks better generally especially during the day.

If your monitor's gamma is 2.2 then VLC is roughly correct, but only on the assumption that most content - especially of that era of television - would have been mastered on 2.4 monitor.

Well, that is interesting to contemplate.. since you were talking about bt.601 etc. recently.. so the question then becomes, why doesn't madvr adjust? Unless it's simply just that public releases stopped before it could be addressed?

VLC's behaviour can be reproduced with madvr's color & gamma settings if you desire. Just keep in mind that room lighting affects perceived gamma due to "simultaneous contrast" effect - in a daylit room 2.2 can look similar to 2.4 in a dim surround.

I turned on enable gamma processing to see, and 2.6/pure power curve looks pretty similar, but could stand to be a bit darker still, 2.7 or more perhaps but it doesn't go past 2.6 as an option.

I know what you mean about adjusting room light affecting perception, I do that all the time to make the picture have more depth, but the thing is, adjusting the lighting isn't going to at-all fix the washed out gamma issue.. if the room has all the lights off, or turned up high, it's still going to be washed, I'm still going to see the macro-blocks on dark scenes, etc. So it actually does need to be right when it is processed.

I don't know what the issue is, if it's the file, if it's madvr, or if it's something else like vlc's d3d11 renderer being bugged and just happened to coincide with being correct for this file? But one thing I do know, this odd gamma issue with bt.601 doesn't happen really with bt.709. Just now I threw up a different file and compared them (with enable gamma turned off). On a scene that I paused, the madvr rendering had slightly better rendering in the darker ground shadow areas, but vlc's d3d11 had slightly better rendering in the highlight areas (clouds), but overall, it wasn't doing the great difference that it was with the bt.601 file.

MPCVR renderer does the same thing with the bt.601/pal file like madvr, washed out gamma.

PS. I haven't watched it yet, but the images were from a 1978 British series by Terry Nation called Blake's 7. "Blake's 7's narrative concerns the exploits of political dissident Roj Blake, who leads a small group of rebels against the forces of the totalitarian Terran Federation that rules the Earth and many colonised planets. The Federation uses mass surveillance, brainwashing and drug pacification to control its citizens."

flossy_cake

19th January 2023, 05:51

More than that.. the colours themselves are different.. if you look at the red and blue in the first row, I guess similar to the same issue you had with Family Guy, except it doesn't have the blue cast. Unless the different shade of red, etc. is due only to gamma? Might be it, more below.

Yep, applying gamma exponent to RGB channels changes the hue and saturation as well. Here's madvr with +0.2 gamma on RGB channels vs VLC:

https://i1.lensdump.com/i/RxH5z7.png

The amount of red in the cyan section is still quite different though - RGB (79,155,181) vs (73,154,182).

The red sections behind it are identical though. Maybe something to do with gamut conversion...

Everything is at 0, and enable gamma processing is unchecked.

So you haven't told madvr that your monitor's gamut is BT.709 ? If not, then gamut will be wrong (assuming your monitor is BT.709). VLC may be assuming monitor is 709 gamut by default.

Sunspark

19th January 2023, 06:43

So you haven't told madvr that your monitor's gamut is BT.709 ? If not, then gamut will be wrong (assuming your monitor is BT.709). VLC may be assuming 709 gamut by default.

I was working on the assumption that if it was on "disable calibration controls" it's supposed to assume BT.709/2.2 and do everything from there in terms of making changes regarding bt.601 content, etc. Played around with it, if I tell it already calibrated to BT.709 the picture doesn't change whether I set it to 1.8 or 2.6 and seems to be the case still whether on or off.

However, turning on only enable gamma processing changes things.

So if I tell it it's calibrated to bt.709/1.8 and set 2.40/pure in enable processing, it looks basically the same between the two players actually!

So, is my understanding correct, that one needs to manually enter in the processing # based on the content? To use 2.4 for bt.601 and to change that to any number between 1.8 and 2.2 for bt.709 content? It would have been nice if it was automatic just based on the "calibrated to" setting..

thighhighs

19th January 2023, 07:01

I would like to know what other users think about 1D. I think 25% AB sometimes helps for video 2160 -> 1080After taking a closer look at SSIM1D100 + AB25% vs SSIM1D75 no-AB (video source), I changed my mind. AB blurs too much. So it's better to use SSIM1D without any AB.

flossy_cake

19th January 2023, 09:30

I was working on the assumption that if it was on "disable calibration controls" it's supposed to assume BT.709/2.2 and do everything from there in terms of making changes regarding bt.601 content, etc. Played around with it, if I tell it already calibrated to BT.709 the picture doesn't change whether I set it to 1.8 or 2.6 and seems to be the case still whether on or off.

However, turning on only enable gamma processing changes things.

So if I tell it it's calibrated to bt.709/1.8 and set 2.40/pure in enable processing, it looks basically the same between the two players actually!

So, is my understanding correct, that one needs to manually enter in the processing # based on the content? To use 2.4 for bt.601 and to change that to any number between 1.8 and 2.2 for bt.709 content? It would have been nice if it was automatic just based on the "calibrated to" setting..

As far as I can tell, in relation to SDR content only, if you don't tell MadVR anything about your monitor in the "this display is already calibrated" section, MadVR won't assume anything about your monitor, and therefore won't modify the colour at all to suit your monitor, and you'll get incorrect colour if the video content uses a different gamut than your monitor's gamut, which is the case when watching SD 601 gamut content on a typical HD 709 gamut monitor or TV. That's what the Family Guy screenshots here (https://forum.doom9.org/showthread.php?p=1980738#post1980738) are showing.

Also in the "this display is already calibrated" section, there is "this display is calibrated to the following transfer function / gamma". This appears to be used by MadVR as part of its gamut conversion formula only. For SDR content it barely affects the midtones (as you have noticed). It's not a gamma control, I think it's just one of the coefficients used in the gamut conversion formula, as far as I can tell.

Separately to this in the other "color & gamma" section is the "desired display gamma / transfer function". This is a real gamma control and does change the entire gamma curve to the value you choose. To do such a conversion requires knowing what your monitor gamma is to begin with. If you haven't told MadVR what your monitor gamma is in the previous section (which happens if you select "disable calibration controls for this display") it assumes your monitor is 2.2. So if you chose 2.4 it would do a +0.2 adjustment. Otherwise, if for example you already told MadVR your monitor is 2.4 in the other section, selecting 2.4 results in no change in colour.

We shouldn't need per-content settings since MadVR automatically gets the source gamut from upstream, and even if upstream doesn't have it, MadVR guesses correctly based on things like resolution and framerate of the video ("best guess" on debug screen) and I've never seen it get it wrong.

So if your monitor has 2.2 gamma and 709 gamut, and you want to get accurate gamut conversion for SD 601 gamut content, and additionally emulate what VLC appears to be doing:

1. calibration > this display is already calibrated > BT.709 & pure power curve 2.2 (informing MadVR your monitor characteristics)
2. color & gamma > enable gamma processing > pure power curve 2.4 (emulating VLC behaviour of assuming SDR content was mastered at 2.4 gamma by applying a +0.2 gamma shift)

The downside to 2 is that it will cause MadVR's HDR->SDR conversion to be 0.2 gamma darker than it should be. To avoid this I only set 1 and not 2, and calibrate my monitor to 2.4 for night and 2.2 for day in the video card gamma table which is system-wide.

After taking a closer look at SSIM1D100 + AB25% vs SSIM1D75 no-AB (video source), I changed my mind. AB blurs too much.

Yeah that's why I stopped using it as well.

Anti ringing seems ok and has a subtle effect when used in conjunction with sharpening filters, but doesn't seem to be effective at removing ringing in the source, eg. S01E01 of The Sopranos Bluray has some ringing that I wanted to get rid of and it couldn't help there at all.

Sunspark

19th January 2023, 10:35

Anti ringing seems ok and has a subtle effect when used in conjunction with sharpening filters, but doesn't seem to be effective at removing ringing in the source, eg. S01E01 of The Sopranos Bluray has some ringing that I wanted to get rid of and it couldn't help there at all.

I vaguely remember madshi saying that the anti-ringing only applies to new ringing created by scaling done by the algorithm and not the original material itself. Makes sense to me, probably easier to write a calculation to compare before and after and see if it made new ringing on edges than to only have 1 image and decide whether it has elements that need to be removed.

Asmodian

19th January 2023, 15:04

probably easier to write a calculation to compare before and after and see if it made new ringing on edges than to only have 1 image and decide whether it has elements that need to be removed.

I don't think it looks for ringing at all, it always runs another algorithm after scaling which makes aliasing a little worse while also making ringing better. It is based on the known amount of ringing added by the scaling algorithm, not an analysis of the image after scaling.

thighhighs

21st January 2023, 17:57

Still testing SSIM algorithms on my videos. This time SSIM2D. I usually use SSIM2D100 + AB100%, somethimes it looks too soft. 75% AB makes it sharper, but when I started using this combination, I noticed that it was actually worse. 2D100 + AB75% looks unnatural on the edges. I made an example, but it's hard to see in the pictures. I hope you can see if you pay attention to the tower of this building. https://imgsli.com/MTQ5MTA5/0/1
Anyway, the only combination that works well for me is SSIM2D100 + AB100%. With other combinations, the result is a strange look.
So, I can't say that SSIM2D is the king.

flossy_cake

22nd January 2023, 06:00

I generally don't feel a need to sharpen native HD sources, but maybe that's because of my small screen size & long viewing distance (51" @ 3.3m).

thighhighs

22nd January 2023, 19:16

Personally, I prefer sharp algorithms for downscaling. Both SSIM or Bicubic150. Other algorithms are far behind. Lanczos is also sharp and I liked it before.

ryrynz

22nd January 2023, 21:06

So, I can't say that SSIM2D is the king.

I'll say it then. It slaps 1D [emoji106]

thighhighs

23rd January 2023, 10:25

Ok :D
I want to say more about the 2D + AB vs no AB. 2D makes a noticeable fatness, I don't like it. It looks like a sharpening filter. So the AB filter is useful, i think.
To back up my words, I made an example. I wanted an example where you can not only notice the difference, but where you can see that the image looks bad. I think this scene fits. It can also be seen that as a result, 2D100AB100 is better than 1D here -> https://slow.pics/c/g7oulciR
edit: Conclusion. As a result of my testing, I decided never to use AB for 1D. And 2D100 + AB100 is not a bad idea, I did not find a better combination.

huhn

23rd January 2023, 14:01

sorry but where do you even look at this source is already super soft it.

no matter how i look at it the difference is superficial in this image. downscale something sharp where ssim can even try to preserve something.

flossy_cake

23rd January 2023, 15:40

It can also be seen that as a result, 2D100AB100 is better than 1D here -> https://slow.pics/c/g7oulciR

I'm really struggling to see a difference... is this some kind of placebo test you are running on us or something :devil:

thighhighs

23rd January 2023, 16:32

Cars, man between cars, rooftops, T. rex, "cartoon" tree on the left, bottom of the image... Is your screen 2160? Or 1080? I believed that it should be noticeable.

I have no idea for another comparison.
In my example, starting at 2D50, the image doesn't look right at all to me. 75 is actually slightly better, but sharper overall. That's why just lowering the strength won't help much.

huhn
What do you think about the SSIM2D? I remember your posts about Biсubiс150, but not about SSIM2D.

ryrynz

23rd January 2023, 20:59

Personal use case.. If you sharpen a thing that already has a lack of detail on edges you can get some obvious sharpening in places. In saying that, I see no oversharpening at all on 2D without AB. Just choose what looks good to you and call it a day.

thighhighs

24th January 2023, 01:20

Just choose what looks good to you and call it a day.Yes, I have decided everything for myself. I just wanted to back up what I said with an example.
Personal use case.. If you sharpen a thing that already has a lack of detail on edges you can get some obvious sharpening in places. In saying that, I see no oversharpening at all on 2D without AB.This is important. Because if there is no negative effect, AB will not look useful (why reduce the sharpness?). I guess a lot of people like 2D as a sharpening filter. Instead, I would like an improvement over 1D, no more. Combining with AB is the best way.

Mb these will be better... I compare only 100, others will only create confusion
Here 2DAB100 has an advantage over 1D. 2D100 is oversharp -> https://slow.pics/c/Wh67f9nQ
Here AB does not allow swelling. Close to 1D but you can find different pixels -> https://imgsli.com/MTQ5ODIw/0/1

flossy_cake

24th January 2023, 03:32

Cars, man between cars, rooftops, T. rex, "cartoon" tree on the left, bottom of the image... Is your screen 2160? Or 1080? I believed that it should be noticeable.

The resolution of my screen shouldn't really matter (2560x1440) as long as I'm at 1:1 pixel mapping. I've tried moving closer/further from the screen, still can't see any appreciable difference.

I *can* just see a small difference in the white hair on your next screenshot. I would call that a JND (just noticeable difference) but only in A-B comparison - no way I'd pick it in a double blind or if the actor's head is moving.

thighhighs

24th January 2023, 10:23

I *can* just see a small difference in the white hair on your next screenshot. I would call that a JND (just noticeable difference) but only in A-B comparison - no way I'd pick it in a double blind or if the actor's head is moving.You didn't say anything about the picture with the wizard in the forest. Take a look at it if you're interested. https://slow.pics/c/zGHO2QKB
The same effect can be seen on the skin (nose, for example).

flossy_cake

24th January 2023, 13:28

You didn't say anything about the picture with the wizard in the forest. Take a look at it if you're interested. https://slow.pics/c/zGHO2QKB

Ok, I think I know what's going on now. My PC monitor has very little or no sharpening of its own, so the difference is way smaller on PC monitor. Whereas on my TV the difference is more noticeable, I'm guessing because my TV has its own sharpening (I have it set conservatively though - 20/100 on Samsung TV video processor scale, which is the default for "Movie" picture mode).

Sunspark

24th January 2023, 21:45

ryrynz

24th January 2023, 23:52

I don't downscale any more as I have a 4K set, but 2D was quite superior to 1D. After my initial review of AB I decided I didn't find it useful, better imo to just reduce the sharpness of the display, but it all depends.That's exactly why I wrote the above. I was going to say it again but I edited it out of the post as I was repeating myself. Your display sharpness should be balanced with any other image processing you're doing.

thighhighs

25th January 2023, 02:18

This forest picture is ridiculous. Even at 400% magnification, the only change I noticed was the black on the rim around the wizard's hat changing in thickness which could just simply be a rendering artifact and maybe, an incredibly subtle fraction of a fraction change in texture of stuff. The fact that I say maybe is how indistinguishable it is.

Don't shave hairs this finely. The power consumption isn't worth it at all especially if you can't see any changes.Just to clarify: and nothing else at all? The fact that there are 3 different algorithms does not bother you?
I don't downscale any more as I have a 4K set, but 2D was quite superior to 1D. After my initial review of AB I decided I didn't find it useful, better imo to just reduce the sharpness of the display, but it all depends.AB changes the result of the algorithm.
For the screen I using the "source as is" setting. But it's not 0. If it does anything, I don't expect much. To "fix" the 2D100, I need to gimped the screen so that I can't see anything at all :D And even then I could see AB/no AB.
Why configure the screen based to SSIM2D at all? If the special patterns are ok, everything is fine.

ryrynz

25th January 2023, 05:28

Stick with whatever works for you, your content and display, that's all that matters.. The differences are close to insignificant.

Magik Mark

25th January 2023, 07:57

Hi!

Would u consider rtx 3000 series with or without LHR? Does it make any difference?

thighhighs

25th January 2023, 12:20

Stick with whatever works for you, your content and display, that's all that matters.. The differences are close to insignificant.But that's not what you meant before. You said (again) the SSIM2D can't look weird, it's just that my screen isn't set up properly or something.
I understand that you liked it and everything looked great for you. So AB doesn't make sense. I can accept it.
This is the only algorithm in madVR with this level of sharpness. This is insane changes. But if you don't see sharpening artifacts, you don't think the result looks weird. There is something similar with NGU Sharp. As you know NGU Sharp has an cartoon-look problem. Does it has sharpening artifacts or does it just look weird? Is this really a problem? Do all users notice this when they watch their videos? Or does it just look great to them? That's all I wanted to say.