Why radius of ewa_* 4tap is 4.241?

[rainlycoris]

How did get this value? I know 3taps is 3.28..., 4taps is 4.2410628637960698819573.
And why many decimal places at blur of ewa_lanczossharp or ewa_lanczos4sharpest? I believe this is value isn't made by human eyes.
What is the effect if I don't apply blur, Like ewa_lanczos4? That is also what I'm curious about. Why isn't have this algorithm in mpv.

[llyyr]

https://code.videolan.org/videolan/libplacebo/-/merge_requests/560#note_408215

[rainlycoris]

Thank you very much! And answer of first question is this value in ewa_* can achieve normal 4taps right? Would you briefly explain the principle? Thanks.

[rainlycoris]

And this code run 3taps obtain 0.88451209326050, dosen't equal 0.98... at ewa_lanczossharp .

[rainlycoris]

And this code run 3taps obtain 0.88451209326050, dosen't equal 0.98... at ewa_lanczossharp .

I got it, this is why sharpest different with sharp.

[rainlycoris]

Say some digressions, If I used ewa_hanning with 4taps, did I make a mistake in math?
My goal is to make it clearer.
It has to be said that this looks like still less ringing.

[rainlycoris]

This is my fantasy, there is no observable difference between 3tap and 4tap, even at the pixel level.

my screenshot was wrong, I discovered that the built-in screenshot of the mpv does not include zooming, the correct screenshot below.

ewa_hanning 3tap

ewa_hanning 4tap ewa_lanzcos

[kasper93]

Can you post exact commands you used? And what is the goal you want to achieve?

[rainlycoris]

Can you post exact commands you used? And what is the goal you want to achieve?

My goal is to make it sharper, And control ringing problem such as used ewa_hanning.

My current problem is whether ewa_hanning 4tap is sharper than 3tap. unfortunately, I didn't see it. Is this my illusion? because 4tap ewa_lanzcos boost significantly than 3tap. 4tap ewa_hanning should be better on math?

I used

scale = ewa_hanning
scale-radius = 3.2383154841662362076499/4.2410628637960698819573 (For the example I gave)
cscale = catmull_rom
dither-depth = 10
correct-downscaling = yes
linear-downscaling = yes
sigmoid-upscaling = yes
deband = yes

I have two more questions for you, by the way. If I used scaler-lut-size = 10（Although I knew now this selection is hard-coded, and value is 8）and fbo-format = rgba32f can be better?

Thank you very much!

[norinoriko]

I'm going to be completely honest, there's no benefit to using 4-lobe ewa_hanning over 3-lobe ewa_lanczossharp. The sharpness difference isn't perceivable but the former rings very badly due to having a more prominent halo. This kind of defeats the purpose of even adding the Hann window in the first place.

There isn't even that much of a benefit to using 4-lobe ewa_lanczos, hence why ewa_lanczos4 will never exist as an aliased option in mpv. The only reason why ewa_lanczos4sharpest even has a usecase is because the 3-lobe "sharpest" variant of the filter is subpar for upscaling and has nasty aliasing for a polar filter. Meanwhile the 4-lobe variant performs better in comparison (though it isn't strictly better than ewa_lanczossharp, it just has different usecases).

What you're actually looking for is either antiringing and/or a sharpening filter, as you'll just get diminishing returns by trying to tweak built-in scalers. Many people have already played around with this for decades (some even made careers revolving around this) with very minimal success. At a certain point you just have to throw in the towel and accept that lanczos is endgame. :)

[rainlycoris]

I think you're absolutely right.

So what should I use? ewa_hanning or ewa_lanczoss with anti-ringing or just lanczos (ewa_lanczos is strictly better that lanczos right?). But mpv handbook said that anit-ringing isn't affect any polar scalers. I'm surprised that 4sharpest does exactly that and used exactly this parameter.

Is there anything else I need to do on scale, such as for chroma? No one seems to have come up with a best chroma or dscale routine algorithm . Of course, in fact, I can't see much difference, I just want to get the best result in math. I'm sure there are others who feel the same way, especially now that computer power is quite adequate. thank you for your help.

[norinoriko]

So what should I use?

ewa_lanczossharp is generally accepted to be a good all-rounder filter, and combining it with antiringing is popular among users.

ewa_lanczos is strictly better that lanczos right?

It has less prominent ringing and less aliasing in exchange for being slightly blurrier than tensor lanczos. Both filters have acceptable uses but ewa_lanczossharp is generally higher quality for most people.

From sharper (more ringing) to blurrier (less ringing):
ewa_lanczos4sharpest > lanczos > ewa_lanczossharp > ewa_hanning

Anti-aliasing (this ranking is the same for hash pattern dampening):
ewa_lanczossharp > ewa_hanning > ewa_lanczos4sharpest > lanczos

But mpv handbook said that anit-ringing isn't affect any polar scalers.

Antiringing only works with polar filters when using --vo=gpu-next. --scale-antiring=0.8 is hardcoded for ewa_lanczos4sharpest because most people agree that it looks bad without antiringing.

No one seems to have come up with a best chroma or dscale routine algorithm

If performance isn't an issue then there's no point in using different filters for --scale and --cscale.

mpv's default --dscale=hermite is already "high quality" but there isn't a perfect solution for downscaling, as downscaling is a balancing act between sharpness and aliasing. The only issue with hermite is that it can exhibit moiré patterns at high downscale factors due to its small radius, but this isn't an issue for most people. You can play around with some polar filters, but these are exceptionally slow when downscaling, and some of the better polar filters require antiringing when scaling in linear light (which just makes the filter even slower).