RGBA image losing its transparency when saved as pdf

[pubpub-zz]

What did you do?

convert an png to image to pdf or JPEG2000
edit:
after checking with various image viewer, I've observed that the issue could be linked with my original viewer(nomacs)
In order to not bother any I will left on side JPEG2000 and carry on my investigations

What did you expect to happen?

have the good alpha channel

What actually happened?

alpha channel not correctly displayed in pdf is not the good one

What are your OS, Python and Pillow versions?

• OS: windows 11
• Python: 3.10
• Pillow: 10.0
  with this png
  

from PIL import Image
img = Image.open("tt.png")
print("mode : ",img.mode)
img.save("tt.pdf")
img.save("tt.jp2")
img = Image.open("tt.jp2")
print("mode jp2 : ",img2.mode)
img.save("tt.jp2")

output
mode : RGBA
mode jp2 : RGBA

but the image in the PDF has lost its transparency
same for the jpeg2000 image, although still reported as RGBA

[homm]

I can't confirm JPEG2000. Your image is transparent after conversion.

[pubpub-zz]

@homm
this is the pdf I'm getting:
tt.pdf
the background is black whereas the page is normally white
can you confirm you have the same result ?

[homm]

This is how your pdf is rendered to me. It's unclear to me if PDF really can be transparent.

[radarhere]

I'm also unable to replicate the problem with either format. Are you sure that you are using Pillow 10.0.0, and not a version before 9.5.0?

[pubpub-zz]

I reinstalled it before to confirm the bug was on the latest version.
Actually it looks like the problem is also linked with the viewer:
mozilla/pdf.js#5649 (comment)

Personally I checked all under windows with acrobat reader (black) sumatra (transparent) pdf.js( black and damaged) chrome (black)

[mnmtz]

I am using Pillow 10.0.0 and getting the same incorrect results using the code above:

Transparent_Image
Transparent_Image.pdf

Transparent_Image2
Transparent_Image2.pdf

[radarhere]

How did you install Pillow?

What version of OpenJPEG are you using?

>>> from PIL import features
>>> features.version_codec("jpg_2000")
'2.5.0'

[mnmtz]

I have installed it using pip on Windows 11. The OpenJPEG version is 2.5.0.

[homm]

Sorry, people, could you be a bit more verbose here? You are claiming that the initial problem is related to both PDF and JP2K. So could we isolate this? For example, if you are sure that the problem with PDF happens because of embedding bad JP2K, let's focus on the JP2K problem here. Otherwise, if there are separate problems, let's consider them separate.

I've tried to convert the example image to JP2K above and experienced no problems, it has transparency. If some of you have different results, could you please at least mention this and provide your result with broken transparency and expose why you think it s broken (software you are using to viewing, for example).

[mnmtz]

The problem is not (only) related to JP2K, because in my example you find 2 transparent png images, both are not correctly converted to PDF.

The PDF files have been opened with:

• Acrobat Reader 2023.003.20244
• Google Chrome 115.0.5790.110
• qpdf 11.5.0 (extracting the image)
• macOS 13.5 Preview App

[homm]

"is not related to JP2K" and "is not only related to JP2K" are two completely different statements.

@pubpub-zz could you provide any example or proof that you are experiencing problems with JP2K?

[homm]

I am using Pillow 10.0.0 and getting the same incorrect results using the code above:

Why do you consider these results incorrect? Which result would be correct?

I've tried to convert the second image to PDF using macOS Preview App (with "Alpha" checkbox applied), the results are identical:

[mnmtz]

Your results are correct, but the PDFs generated using Pillow 10.0.0 look like this (on all PDF viewers mentioned above):

[homm]

but the PDFs generated using Pillow 10.0.0

The file opened in the right window on my screenshot is your PDF from the comment which claimed as generated using Pillow 10.0.

[pubpub-zz]

My test showed that the viewer allows to show or not the transparency. I have the same feeling about jpeg 2000 viewer (the image was not good with nomacs but displayed correctly in gimp and imagemagick)

digging further in the pdf specification I found this:

I've removed ColorSpace to take into account
I've set the SMaskIn to 1 and the transparency "appeared" 😁 in Acrobat Reader.
I've checked in pdf.js and there the image is damaged:

removing the "BitsPerComponent fixed this:

There is still a color inversion but I think this is a bug in pdf.js(mozilla/pdf.js#16782)

to sum-up, can you :

• add SMaskIn=1 (to enable alpha decoding)
• remove ColorSpace and BitsPerComponents (to allow the extraction of data from the image)

with the fixed PDF:
tt1.pdf

[radarhere]

I've created PR #7316 to set SMaskInData to 1.

I've removed ColorSpace to take into account

Could you explain this some more?

remove ColorSpace and BitsPerComponents (to allow the extraction of data from the image)

What makes you think that the color inversion is a bug in pdf.js, but not the BitsPerComponent problem? Does "allow the extraction of data from the image" mean something special, or is that just another way of saying "appear correctly in a viewer"?

[pubpub-zz]

I've created PR #7316 to set SMaskInData to 1.

👍

I've removed ColorSpace to take into account
Could you explain this some more?

This is an extract from Pdf 1.7 reference:

(...)

For the same reason we should ignore BitsPerComponents

remove ColorSpace and BitsPerComponents (to allow the extraction of data from the image)

What makes you think that the color inversion is a bug in pdf.js, but not the BitsPerComponent problem? Does "allow the extraction of data from the image" mean something special, or is that just another way of saying "appear correctly in a viewer"?

I've tested a few set of PdfViewer( Acrobat / Sumatra / pdf.js / chrome) and the issue was observable in PDF.js. previously this time of error was linked with Decode entry but setting the Decode entry was not improving that.

[radarhere]

Thanks. I've updated the PR to remove those properties.

I presume that is sufficient to fix the original problem here. 'Transparent_Image2' is also an RGBA image, so I think it should fix that as well.

[radarhere]

'Transparent_Image' is different, because it is not RGBA, but P.

I've created PR #7323. With it, running

from PIL import Image
img = Image.open("transparent_image.png")
img.save("transparent_image.pdf")

gives
transparent_image.pdf

This should clearly be an improvement, because the background is no longer green, but @mnmtz, let us know if you think that is fixed.

If that is all good, then everything in this issue should be resolved.

[mnmtz]

As for me, it looks like that the PR is doing the trick.

Thank you!