[Regression] image save() corrupted the image color start from Pillow 8.3.0

[yusharon]

What did you do?

import io
import numpy as np
from PIL import Image

#generate random data with 21 classes
num_classes = 21
height = 200
width = 300
data = np.random.rand(num_classes, height, width)

#prepare and generate image from numpy array use mode P
pred = np.squeeze(data)
pred = np.argmax(pred, axis=0)
img_with_correct_colors = Image.fromarray(pred.astype('uint8'), mode='P')
print(img_with_correct_colors.getcolors())

#save image uses BytesIO
fp = io.BytesIO()
img_with_correct_colors.save(fp, 'png')
img_with_wrong_colors = Image.open(io.BytesIO(fp.getvalue()))
img_with_wrong_colors.getcolors()

What did you expect to happen?

I expect image contains correct number of class before and after saving (which should have 21 number of classes)

What actually happened?

I run the exact same commands on Pillow 8.3.0 and Pillow 8.2.0

Pillow 8.2.0 gives the correct result

The img_with_correct_colors produce correct number of class

Image before saving(container 21 classes):

>>> print(img_with_correct_colors.getcolors())
[(2883, 0), (2847, 1), (2852, 2), (2889, 3), (2850, 4), (2870, 5), (2826, 6), (2846, 7), (2861, 8), (2836, 9), (2764, 10), (2915, 11), (2831, 12), (2852, 13), (2873, 14), (2826, 15), (2880, 16), (2866, 17), (2756, 18), (2928, 19), (2949, 20)]

After save, the image saved preserves the correct number of class as well (container 21 classes):

>>> img_with_wrong_colors.getcolors()
[(2883, 0), (2847, 1), (2852, 2), (2889, 3), (2850, 4), (2870, 5), (2826, 6), (2846, 7), (2861, 8), (2836, 9), (2764, 10), (2915, 11), (2831, 12), (2852, 13), (2873, 14), (2826, 15), (2880, 16), (2866, 17), (2756, 18), (2928, 19), (2949, 20)]

However, starting with Pillow >= 8.3.0 (Also tested on latest Pillow version 9.1.0)

Image before saving is still correct but image after saving produces wrong number of colors. The result looks like except the class 0 everything else got aggregate to the class 1.

Image before saving:

>>> print(img_with_correct_colors.getcolors())
[(2846, 0), (2940, 1), (2861, 2), (2834, 3), (2886, 4), (2911, 5), (2739, 6), (2797, 7), (2751, 8), (2960, 9), (2900, 10), (2880, 11), (2893, 12), (2810, 13), (2851, 14), (2870, 15), (2780, 16), (2775, 17), (2863, 18), (2954, 19), (2899, 20)]

Image after saving:

>>> print(img_with_wrong_colors.getcolors())
[(2846, 0), (57154, 1)]

What are your OS, Python and Pillow versions?

• OS: Mac OS
• Python: Python 3.7
• Pillow: Pillow 8.2.0 / Pillow 8.3.0 / Pillow 9.1.0

[hugovk]

Bisecting between 8.2.0 and 8.3.0 points to 4d36fee from PR #5552.

[radarhere]

If you would like an immediate fix, I've found the following options work -

1. Specify the number of colors when saving to PNG - img_with_correct_colors.save(fp, 'png', bits=8).
2. Change the mode to "L". That may not work for the rest of your scenario, but it fixes the example you've provided here.

[yusharon]

Thanks for providing the immediate workaround! Going to try and see if that works.

[yusharon]

The workaround works! Thanks for the help! Curious about what change causing this issue and do you have any ETA for the fix. Thanks!

[radarhere]

Just to be clear - I meant that you could do either 1 or 2 as a workaround, you don't have to do both.

[yusharon]

Yes, we did option 1 as option 2 doesn't apply to us. Thanks for clarify though.

[radarhere]

I've created PR #6283 to resolve this.

The next Pillow release is scheduled for July 1, so if the PR is merged before that, this will be fixed then.