getbbox() returns bounding box on all-black RGB image

[ghost]

What did you do?

Running getbbox() on this all black RGB image:

Context / Where is this image coming from?

I'm upgrading a project that was heavily affected by the getbbox() alpha channel behavior change (#4849). For example before the change we were doing something like this (the simplest example):

def images_equal(im1: Image, im2: Image):
  return ImageChops.difference(im1, im2).getbbox() is None

Because when comparing some images, the diff is only in the non-alpha channels and so it broke after the 7.1.0 update.

To fix this, we converted these cases to variations of this:

def images_equal(im1: Image, im2: Image):
  return ImageChops.difference(im1, im2).convert('RGB').getbbox() is None

This works in 99% of the cases, however we have one test case where it isn't because Pillow, for reasons I can't figure out, returns this (rogue ?) bounding box on this all black diff image.

What did you expect to happen?

I expected None to be returned.

What actually happened?

Pillow returned bounding box (8, 7, 995, 223).

What are your OS, Python and Pillow versions?

• OS: Linux 5.8.0-7630-generic #32~1609193707~20.10~781bb80-Ubuntu SMP Tue Jan 5 21:29:56 UTC 2
• Python: Python 3.6.12
• Pillow: 8.1.0

Repro

from PIL import Image

image = Image.open('./weird_bbox.bmp')
print(image.getbbox())

[nulano]

Adjusting the levels in an image editor, I can see that the image is not completely black:

[ghost]

Wow. I even iterated over all the pixels and checked that they are (0, 0, 0) 🤔 Maybe I made a mistake there. Maybe after all my images (see context detail box) aren't the same... I'll check and maybe close.

The question then kinda becomes, why was this not detected in RGBA space, before the change in 7.1.0?

[hugovk]

Adding in the version:

from PIL import Image

print("Pillow:", Image.__version__)
image = Image.open('./weird_bbox.bmp')
print(image.getbbox())

I get the same result for 7.0.0, 7.1.0 and 8.1.0 (macOS, Python 3.6.12):

Pillow: 7.0.0
(8, 7, 995, 223)

Pillow: 7.1.0
(8, 7, 995, 223)

Pillow: 8.1.0
(8, 7, 995, 223)

[ghost]

@hugovk Yes, this image for sure. But before upgrading I was comparing images without converting to RGB, because getbbox behaved differently.

Anyway, after sleeping over this one more night I finally think I figured it out. Something else about the upgrade, made the resulting image look slightly different from the fixture created on 7.0.0. I just got tricked into attributing it to the getbbox change because it was the cause for 99% of the other test failures.

I'm sorry for wasting time everyone. I was just brain-fried by the getbbox breaking change. Thanks for your help.