8.3.0: convert on a P image with transparency now returns a RGB image (used to be RGBA)

[benjaminrigaud]

What did you do?

I converted an image in P mode with transparency on the new 8.3.0 release.

What did you expect to happen?

I expected the output to be RGBA.

What actually happened?

it is now in RGB.

What are your OS, Python and Pillow versions?

• OS: Ubuntu 18.04
• Python: 3.7
• Pillow: 8.3.0 or 8.3.1

Please find the code and image to reproduce the issue on https://gist.github.com/benjaminrigaud/d98a4b54fb409bc8f745d169df0196fc

def test_convert_p():
    with open("indexed_with_transparency.png", "rb") as f:
        data = BytesIO(f.read())

    img = Image.open(data)

    assert img.mode == "P"
    assert img.palette.mode == "RGB"
    assert img.info.get("transparency") == 0

    img2 = img.convert()

    assert img2.mode == "RGBA"

[radarhere]

This would be due to 804183c. It changed palettes for P images with info["transparency"] to RGB instead of RGBA. Separate to that, converting from a P image with no mode argument uses the palette mode if there is one.

I've created PR #5594 to fix this.

[NicolasHug]

I am facing a similar issue when converting a P image into LA: values of the transparency channel are all 255 with pillow>=8.3:

from PIL import Image
import numpy as np

# same image that OP mentioned https://gist.github.com/benjaminrigaud/d98a4b54fb409bc8f745d169df0196fc
path = the_path_here

with Image.open(path) as img:
    img = img.convert("LA")
np.unique(np.asarray(img)[:, :, -1])

This will print array([255], dtype=uint8) on 8.3.1 and array([ 0, 255], dtype=uint8) on 8.2

Using another P image like https://github.com/pytorch/vision/blob/master/test/assets/fakedata/logos/palette_pytorch.png, the results are even more drastic:

array([255], dtype=uint8) vs array([ 0, 16, 64, 80, 96, 128, 160, 208, 224, 240, 255], dtype=uint8)

[radarhere]

I've created #5606 to fix this second problem. I validated the result just by visually inspecting the output.