What is even the point of Blackman resampling?

[Katie Boundary]

To achieve any given amount of sharpness, it requires ~50% more taps and CPU cycles and causes ~50% more ringing than Lanczos, for no apparent benefit :confused:

[Dogway]

In my tests it achieves same sharpness with a tiny bit less ringing at the cost of performance (more taps), but YMMV, maybe DTL can give a more detailed read of the kernel.

[Katie Boundary]

same sharpness... less ringing... more taps

That's... mathematically impossible with windowed sinc filters? More taps equals more rings, so if you hold ring depth (sharpness) the same, then more taps equals more ringing, and less ringing equals fewer taps. Conversely, more taps/rings with less ringing requires shallower rings and therefore less sharpness.

[Dogway]

Yes, more taps involves more ringing within the same sinc filter, but blackman window is more tapered, see graph here (https://legacy.imagemagick.org/Usage/filter/graph_windowing.jpg). In either case it's splitting hairs, as the difference is not by much (Blackman6 vs Lanczos4), kind of placebo.

[Katie Boundary]

Yes, more taps involves more ringing within the same sinc filter, but blackman window is more tapered, see graph here (https://legacy.imagemagick.org/Usage/filter/graph_windowing.jpg).

I've seen it. It's not the graph that you should be looking at. Look at the filter shapes. For any Blackman filter of N lobes, a Lanczos filter of 2N/3 lobes will almost perfectly reconstruct the first 2N/3 lobes of that Blackman filter. All Blackman does is add 50% more lobes that are too shallow to make a visible difference in the final output. (Black is raw sinc BTW)

https://i.imgur.com/PtwxrLA.png

[Katie Boundary]

In fact, Lanczos-2 and Blackman-3 are even better approximations of each other than Lanczos-2 and Catrom (purple) are. Spline16 was unavailable for comment.

https://i.imgur.com/cPXKG5s.png

[Dogway]

With Blackman you get the same (or better) sharpness of lanczos at less ringing because the ripple spread. The difference is minimum though, and you have to consider this is on gamma encoded imagery.

https://i.imgur.com/MCCvg6O.png https://i.imgur.com/KnISU23.png

Here you can see how Blackman has a slightly increase in texture sharpness, and on edges sometimes the 2nd negative ring is more spread. If it's worth the two more taps CPU cycles that's another thing.

[Katie Boundary]

https://i.imgur.com/M8zHm9k.png

[Dogway]

They are not, but I understand that for some people it would be considered placebo territory.
https://i.imgur.com/Ftd5R8K.png

[StainlessS]

Agree,

L=Imagesource(".\MCCvg6O.png",end=0)
R=Imagesource(".\KnISU23.png",end=0)
Subtract(L,R)
Z=2
Levels(128-Z,1.0,128+Z,0,255)

https://i.postimg.cc/Qdg8Hr7C/KB.jpg (https://postimages.org/)

[Katie Boundary]

Agree,

With which part?

[StainlessS]

All parts.
Katie, chill man, life aint that long.

[DTL]

Resizers in AVS are not clearly named. Some are only different versions of SincRezise with differently weighted sinc kernels. These are BlackmanResize (Blackman window), LanczosResize (Lanczos window), SincLin2Resize (sort of trapezoidal window) and SincResize (no or 'rectangular' window). Others use significantly different shapes of kernels.

Difference between Lanczos and Blackman window may be important for some DSP applications (in the form of Fourier spectrum shape of these windows - https://en.wikipedia.org/wiki/Window_function ) but may now significantly visible when these weightings used in image data resizing. With a big enough number of taps all sinc-based resizers may look very close (because starting lobes of sinc of max amplitude become close to unweighted).