Brightness clipped on I;16 images loaded from png

[flagadum]

What did you do?

• Open a 16bit greyscale image from a png file
• Paste that image into another 16bit greyscale image (I;16)

What did you expect to happen?

The pasted image should be kept the same

What actually happened?

The bright pixels got clipped to middle grey (32767 ?)

Proposed workaround?

Immediately after loading png, convert the image to array, then back to image again.

What are your OS, Python and Pillow versions?

• OS: MacOS 10.15
• Python: 3.9.10
• Pillow: 9.0.0

from PIL import Image
import numpy

# Let's create a white 16bit square 
img1 = Image.new("I;16",(256,256),(65535))

# Let's save this square as png and open it again
if True :    # if False: The bug does not happen
	img1.save("test.png")
	img1 = Image.open("test.png")

if False :    # If True: Bug workaround: Convert image to array and back to image again
	(sizeu,sizev) = img1.size
	img1 = Image.fromarray(numpy.array(img1.getdata()).astype(numpy.uint16).reshape((sizev,sizeu)))

# We check that the loaded square is white
img1.show()

# But we might already see that something is wrong here
print(img1.histogram())

# Let's paste this square in another 16bit image
img2 = Image.new("I;16",(288,288),(0))
img2.paste(img1,(16,16,272,272))

# Is the pasted square white or grey???
img2.show()

[radarhere]

Currently, Pillow opens I;16 PNGs as I. See #3041 or #3796

from PIL import Image

im = Image.new("I;16", (1, 1))
im.save("test.png")
reloaded = Image.open("test.png")
print(reloaded.mode)  # I

So a workaround would be to save it as a TIFF instead.

[radarhere]

Another workaround would be to create img2 as "I" rather than "I;16".

[radarhere]

I've created PR #6112 to resolve this particular scenario, by fixing incorrect clipping when the "I" mode PNG was converted to match the "I;16" image you were pasting it on.

However, Pillow will still clip the values of the 32-bit "I" mode when converting to the 16-bit "I;16" mode. This scenario is only fixed by my PR because you are using white. If you try this with other images, I expect you will find a problem.

So resolving this situation more generally can be either part of #3796 or #3159.