What is the difference between Lanczos seperable and non-seperable.

[aitte]

Irident Developer is doing the right thing!

 

Lanczos2 (smooth) = separable, 2 lobes

Lanczos3 (sharp) = separable, 3 lobes

Lanczos5 (sharper) = separable, 5 lobes

 

The number of lobes = the sharpness. Common values are 2 (almost zero ringing artifacts) and 4 (best tradeoff between sharpness and ringing).

 

Again, AFP is misimplementing Lanczos by having a pointless, slow non-separable mode and by not providing varying lobe (sharpness) settings.

 

I wish we had Irident developer's menu. It's like an A-Z of all good scaling algorithms!

[Magnus_vb]

Is there a set of pictures that demonstrate these differences?

 

Sorry. No images. But the help file have a link to these (technical) examples

[aitte]

I posted a link to image examples for all common filters earlier in the thread.

[peter]

Thanks for the lighthouse and single pixel image comparisons.

 

 

Whenever I scan newspapers or magazines: I end up with a moire pattern. However a quick fix, is to use a very low amount of Gaussian Blur. That generally does the trick. 

 

Are there any other tricks like this?..faulty image/name of filter/amount of filter applied.

[gary1948]

Ok I'm just a poor graphic designer, up in years, who has used PS for color correction, resizing and setting up to print ready for a web press.

 

Can anyone compare the Affinity settings to the ones in PS?

 

That's all I need...... Don't care about x and y. Never much cared for algebra anyway

[JDW]

Can anyone compare the Affinity settings to the ones in PS?

That's all I need...... 

 

That would be all most of us need!  :-)

 

But no, I've not seen a comprehensive comparison on a single chart.  I myself have just been trudging through the Affinity apps trying to compare what is what.  Thankfully, some keyboard shortcuts are the same, but not all.  Although I bought Designer and Photo, neither app is still an Illustrator and Photoshop replacement yet.  I am hopeful that day will come and come soon though.  And toward that end I am searching the apps to see what can be improved.  We should all remember that Photoshop and Illustrator have been around FOR DECADES and are very refined apps.  That makes the Affinity apps all the more surprising because they are really good after only 5 years of engineering.  Perhaps in another year or two, if we all submit well thought out feature requests, we can retire Adobe for good in 2016 or 2017.  Fingers crossed.

[Figmatt]

Since no one knows what the difference is between Lanczos 3 "separable" and "non-separable" may I suggest that Serif put something more meaningful and descriptive into those parenthesis instead?

 

In my testing, Lanczos 3 (separable) enlargements and reductions a tad sharper than Bucubic, and Lanczos 3 (non-separable) is much sharper.

 

So I would think the Resample popup should read as follows:

• Nearest Neighbor
• Bilinear
• Bicubic
• Lanczos 3 (sharp)
• Lanczos 3 (sharper)

Because, quite honestly, "separable" and "non-separable" may appeal to the left-brainers, but isn't Affinity Photo appealing to our right-brain creative side?  Yes.  So make the descriptions more understandable, please!

Agreed

[harrym]

^^

from all tests I've done I think it's more like

• Lanczos 3 (sharp)
• Lanczos 3 (too sharp)

 

;) 

[Ian Ollmann]

The difference is that separable 2-D Lanczos doesn't exist. We use it anyway. Allow me to explain. 

As has no doubt been explained above, if we can represent a 2D convolution filter as the outer product of two 1-D convolution filters, we can achieve the same outcome of the 2D convolution filter by doing two 1-D convolutions.  Since the number of multiplies (and adds) in a 2-D convolution is proportional to the height x width of the filter, and the number of multiplies (and adds) for doing two 1-D convolutions is proportional to height + width, for heights and widths that are large enough, it is much faster to do the two pass algorithm.

This would of course require that the 2-D lanczos filter be separable into 1-D convolutions.  It isn't.  If we imagine the usual lanczos 1-D waveform distributed radially about a point (this looks a little bit like a dumbbell boat anchor with a large hump at 0,0 and a lower circular depression and speed bump around it), that is a 2-D lanczos resampling filter.  What image processing filters actually do is two 1-D lanczos filters. Notably when you multiply two 1-D lanczos filters together you don't get the 2D radial lanczos filter, you get something else. Consequently, the "separable" lanczos delivers pretty loud ringing aligned to the horizontal and vertical axes of the image. If we used a true 2-D lanczos filter, then those would tend to cancel out, and you'd get more circular ringing around bright or dark points. I'll assert that the 2D filter probably looks better but reasonable people may disagree, it probably depends on the image and one clearly is much slower. 

Given the relationships between convolutions and FFT, there is probably a somewhat speedy-ish solution to this problem in the frequency domain for large enough resampling filters, but given the size quantization issues behind FFT algorithms, the fact that we are downsampling and the need to support arbitrary downsampling ratios there is inevitably a hack such as a fast gaussian blur downsample that times in better and gets used instead. Some people even prefer gaussians because they don't ring, though they can be a little blurry.

[Figmatt]

Amazing that every time I choose this option I get stopped in my tracks, paralysed by the choice of "separable" or "non-separable"

🤪