Artifacts when scaling down with Sinc Lanczos3

[Jason Gerard DeRose]

Binary package hint: gimp

Scaling down a solid white field with the Sinc algorithm causes a weird thatched-pattern to appear. I've seen this before but now have an easy way to reproduce. I'm not sure whether the resolutions matter, but here's how I'm currently reproducing:

1. Create a new RGB image filled with solid white, size 5616x3744 (size of 5D Mark II images, what I was editing when I found it)
2. Scale down to 1200x800 using "Sinc (Lanczos3)" Interpolation
3. There will now be a weird thatched-pattern in the image, even though it should be completely #ffffff pixels

I attached the resulting image, saved as a PNG.

ProblemType: Bug
DistroRelease: Ubuntu 10.10
Package: gimp 2.6.10-1ubuntu3
ProcVersionSignature: Ubuntu 2.6.35-22.33-generic 2.6.35.4
Uname: Linux 2.6.35-22-generic x86_64
Architecture: amd64
CheckboxSubmission: fdbdfcded0c0bb479a6b52e9ec5af131
CheckboxSystem: edda5d4f616ca792bf437989cb597002
Date: Sun Sep 26 14:32:20 2010
EcryptfsInUse: Yes
InstallationMedia: Ubuntu 10.04 LTS "Lucid Lynx" - Release amd64 (20100429)
ProcEnviron:
 PATH=(custom, user)
 LANG=en_US.UTF-8
 SHELL=/bin/bash
SourcePackage: gimp

[Jason Gerard DeRose (jderose)]

Okay, definitely depends on resolutions, specifically the ratio of original to final size.

For example, I created a white-filled 1600x1600 RGB image and tried scaling to a few different sizes:

1/4: 1600x1600 => 400x400 Okay
1/5: 1600x1600 => 320x320 Artifacts
1/8: 1600x1600 => 200x200 Okay

In my original image I was scaling each side to approximately 1/4.7 the size. So it seems that it's at least around this ratio that the bug appears.

[Harenome Ranaivoarivony Razanajato (kshegzyaj)]

I can confirm that resizing at 1/5 of the original size can cause the appearance of some artifacts.

I have tried with the sizes you have been using, and with an original size of 1000x1000.
I have also tried with a 1200x1200 picture resized to 200x200, and the same problem occurs. Maybe the range is wider than you thought, as it also happens around a 1/6 ratio.

Also, it seem to occur only with white. I haven't been able to reproduce the bug if the image is filled with another color.

[Izo (shakhtar9)]

I can't reproduce this at all.

[Harenome Ranaivoarivony Razanajato (kshegzyaj)]

I forgot to say this: for my screen is not perfect, I couln't directly notice the "weird thatched-pattern".
In order to reveal the artifacts, I set the bucket-fill tool's threshold at 0 and then randomly clicked the area. If the scaling went fine, the whole image should be filled. When resizing at 1/5 and 1/6, only bits of the picture or the pattern itself would be filled.

So, in my case, I was not able to see the artifacts because of my screen's quality and settings, but these artifacts did exist.

[Mark Harrison (mark-z-harrison)]

Given the symptoms:

1) Occurs during resizing using factors that are not powers of 2 (i.e., 2, 4, 8, etc.). Comment #2
2) The thatched pattern is the color (254, 254, 254), where pure white is (255, 255, 255). Attachment to Comment #1
3) Only occurs for pure white (255, 255, 255) starting image. Comment #3

I would strongly suspect that this is not a bug, but a rounding error inherent in filter. It just happens that only images starting from pure white end up with rounding errors large enough to change a pixel value by one. The Lanczos filter actually takes on negative values at places, so that might be enough to put a resampled pixel one value below where it should be when the scaling isn't a nice power of 2.

In other words, this is a limitation of the filter, not a programming bug.

[Thomas Hotz (thotz-deactivatedaccount)]

With GIMP 2.8.4 it's not reproduce able for me. Please try to reproduce this with this GIMP version I used. Thank you very much!

[Thomas Hotz (thotz-deactivatedaccount)]

And for the next version (2.10) it is planned, that "Sinc (Lanczos3)" Interpolation will be replaced by a new one that works together with GEGL.

Look here please: http://www.gimpusers.com/tutorials/whats-new-in-gimp-2-10

I'm setting this bug due the link to confirmed.