ImageFilter.GaussianBlur seems broken with numpy>2.0.0 and Pillow>10.3.0

[MVKozlov]

img = Image.open(bytesio_source).convert('L')
img.filter(ImageFilter.GaussianBlur(radius = size))

File "/home/test/recognizer/m_recognize.py", line 47, in im_blended
  img3 = img.filter(ImageFilter.GaussianBlur(radius = size))
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/test/.venv/lib/python3.12/site-packages/PIL/Image.py", line 1375, in filter
  return self._new(filter.filter(self.im))
                   ^^^^^^^^^^^^^^^^^^^^^^
File "/home/test/.venv/lib/python3.12/site-packages/PIL/ImageFilter.py", line 191, in filter
  if xy == (0, 0):
     ^^^^^^^^^^^^

It looks like #8195

Python 3.12, numpy 2.2.4
Pillow 10.3.0 work ok
10.4.0 and 11 - does not

[radarhere]

If I try the following code

img = Image.open("Tests/images/hopper.png").convert("L")
size = numpy.array([3, 7])
img.filter(ImageFilter.GaussianBlur(radius = size))

with Pillow 10.3.0 and NumPy 2.2.4, I still get

ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

So I'm not able to replicate it working with Pillow 10.3.0.

In that other issue, I asked

Is there any particular reason that it is helpful for resize() to accept a NumPy array? I have to imagine that most of Pillow doesn't work well with NumPy arrays as sizes.

So in this situation, is there any particular reason it's not very easy for you to replace

img.filter(ImageFilter.GaussianBlur(radius = size))

with

img.filter(ImageFilter.GaussianBlur(radius = tuple(size)))

?

[MVKozlov]

Sorry, I wasn't very precise. Size in this case is int.
for ex size=2
only the error itself looks like mentioned issue

[radarhere]

Hi. I think there is still a detail missing.

img = Image.open("Tests/images/hopper.png").convert("L")
size = 2
img.filter(ImageFilter.GaussianBlur(radius = size))

passes with Pillow 11.2.1

[MVKozlov]

Sorry again,
this was part of a larger complex procedure, so I couldn't immediately understand the variable types and extract the minimal piece of code that correctly demonstrates the problem.

The problem actually looks like this:

import numpy as np
from PIL import Image, ImageFilter

size = np.int64(2)
img = Image.open("photo.jpg").convert("L")
img.filter(ImageFilter.GaussianBlur(radius = size))

[radarhere]

Thanks. So I can replicate this now. It is a result of #8156.

However, that still leads more or less back to my question - is there any particular reason it is helpful for Pillow to accept a NumPy integer here? I expect there are many more places in Pillow where a regular integer can't be simply be replaced with a NumPy integer, and even more so in the Python ecosystem as a whole.

You could just replace

img.filter(ImageFilter.GaussianBlur(radius = size))

with

img.filter(ImageFilter.GaussianBlur(radius = int(size)))

[MVKozlov]

Yes, when I understood the source of the problem, I solved it exactly like that :)
My issue, perhaps, transformed into "why did it work before, but now it doesn't"

Ok, I think this topic may still be useful for someone