A notable discrepancy:
image
("pillow (linear)" is applying the resize after a manual conversion to linear sRGB space, "reference (sRGB)" is the reference resizer run with --no-gamma-handling.)

The correct behavior would be to convert images to a linear color space before resizing. If the situation has changed at all since #1604 was filed it may be worth revisiting, but in either case it's out of scope for this. As you can see, it matches the reference implementation without gamma handling by default, and when applied in linear sRGB space it matches the reference with gamma handling, both within reasonable quantization error.

Just for reference, this is the amount of difference this produces on a test image:
image

(削除) Another discrepancy I am noticing is a number of spots with relatively large differences along the edges of images, this might indicate that there's a problem with conv padding. Comparisons of Lanczos as implemented in MKS's reference code produces some artifacts like this too, but the artifacts are lower magnitude. This is out of scope for this PR but it needs to be checked on. (削除ここまで)

After some more experimentation (in the form of doing a separate basic implementation of the whole resizing process), I've determined that this is most likely just due to implementation differences. MKS's reference code does the resize as three separate kernels, but here they are composed into a single kernel, and that causes boundary differences. (削除) My understanding is that neither is necessarily more visually correct than the other, and despite some of the pixel differences being fairly high magnitude they are still sub perceptual. (削除ここまで) edit 3/18/25: Further empirical testing has shown me that single kernel implementation is generally more accurate when it comes to handling edges. With that, I am completely satisfied with the correctness of the implementation itself.

(if you're having trouble seeing it, they're especially prominent on the top of the image)
image

I have updated this to include Magic Kernel Sharp 2013 (at the suggestion of the creator -- it should have the same effects in combatting moire/blockiness while being slightly faster than Lanczos 3) as well as the 2021 implementation.

I have also added these kernels to documentation at appropriate points and added them to tests. The additions to concepts.rst may need revision. I have also not added the test_reduce_x and test_enlarge_x tests because I am not completely certain of how those tests need to be properly set up and would want some guidance on that first -- as far as I am aware the current implementation right now is correct and could be trusted to generate testing code for this. The only other point in the code that I think might need some advisement is whether MKS2021 should replace LANCZOS at src/PIL/IcoImagePlugin.py:77. In any case, I'm opening the issue for review.

any thoughts on how this will proceed to be released?