SincResize?

[10L23r]

From the wiki article:

SincResize is added in v2.6 and it uses the sinc function as resizer. See LanczosResize for an explanation of the taps argument (default: taps=3; 1<=taps<=20), however there's only one lobe.

Does anyone know how it preforms compared to the other resizers?

When will 2.6 come out? I can't find it on sourceforge.

[MfA]

No idea how it performs, but that description is wrong. It's simply truncated sinc, so it has as many lobes as taps (ignoring mirroring around the origin).

[*.mp4 guy]

Sinc should have more ringing then lanczos, but be sharper. Sinc should be the sharpest and most highly ringing linear interpolation algorithm for any given number of lobes/taps. Before anyone says that the sharpness will be artificial, merely a result of the halos; remeber that the same (flawed) argument could be used in favor of bilinear scaling, yet you will never get the same detail out of bilinear as bicubic, and bicubic will never match lanczos3, the same holds true for sinc.

[10L23r]

No idea how it performs, but that description is wrong. It's simply truncated sinc, so it has as many lobes as taps (ignoring mirroring around the origin).

That's what I thought. I was confused by that description.

Since the function is truncated, would it be possible to eliminate ringing artifacts by setting taps to 2 (or whatever number is equivalent to taps=2 in lanczos)?

[shoopdabloop]

so would SincResize cause more or less halos than BicubicResize?

[neuron2]

Can't you just try it and see for yourself?

[10L23r]

so would SincResize cause more or less halos than BicubicResize?

I think SincResize would create more ringing and haloing (ant thus sharpening)

neuron2: Where can we get avisynth 2.6? The closest thing I've found is a pre-alpha build from 2006 (http://forum.doom9.org/showthread.php?t=114643), and it doesn't seem to have sinc resize.

[neuron2]

I don't know. Please PM Wilbert about it.

[kemuri-_9]

i compiled a build of avs 2.6 back in may, here's what

version()
SincResize(Width()*2,Height()*2)

looks like:
http://kemuri9.net/forumpics/sinc_resize.png

- as you can see it's riddled with haloing
much more haloing than that of lanczos4 or others.

[shoopdabloop]

..ouch.

[Dark Shikari]

Combine it with dehalo_alpha? ;)

[shoopdabloop]

i don't even think dehalo_alpha could fix that haloing on live action sources without ruining a whole bunch of detail.

[10L23r]

Whoa... that looks disgusting. The inside of the text looks decent though. I think setting taps=2 should work better than taps=4 + dehaloing.

[MfA]

Sinc should have more ringing then lanczos, but be sharper. Sinc should be the sharpest and most highly ringing linear interpolation algorithm for any given number of lobes/taps.
It's easy to make a "sharper" linear interpolation filter, just increase the magnitude of the first negative lobe.

Wouldn't it be better to make the default taps for Sinc 2? At least it's slightly useful that way.

[*.mp4 guy]

It's easy to make a "sharper" linear interpolation filter, just increase the magnitude of the first negative lobe.

Wouldn't it be better to make the default taps for Sinc 2? At least it's slightly useful that way.

No, increasing the relative size of the first negative lobe is not all you need to do to make a sharper linear interpolator. Thats equivelent to messing withn the ringing parameter in bicubic, You can't get the same frequency response from 2 taps as 4, which is why lanczos 4 looks so much sharper then bicubic no matter what you do with it.

The extreme amount of ringing in this case is because of how sharp and aliased the input is. Sinc is more useful on more aggressively lowpassed material. Its the ideal filter for upscaling analog signals, like the horizontal dimension of a television signal, in this situation it should theoretically introduce no ringing and be the most accurate.

[2Bdecided]

You only get "no" ringing if there are "no" frequencies "near" the filter cut-off frequency.

A sinc filter is accurate in terms of frequency response, but it's not much use for images at all.

Cheers,
David.

[Wilbert]

No idea how it performs, but that description is wrong. It's simply truncated sinc, so it has as many lobes as taps (ignoring mirroring around the origin).
I noticed that someone corrected the documentation. Thanks!