How to override jpeg_factory from JpegImagePlugin

[cmahnke]

Currently it's not possible to load a 'raw' as in unmodified JPEG file if Pillow suspects a file is a MPO, this is due to the fact that jpeg_factory in JpegImagePlugin contains a switch if it might be a MPO / UltraHDR file, is there a way to disable these checks
and work with those files?

Jus reregistring a open handler with my own function like this won't work:

Image.register_open(JpegImageFile.format, jpeg_factory, JpegImagePlugin._accept)

Manipulating the registry (except adding things) isn't exposed via API either.

BTW: Detecting images produced by libultrahdr doesn't seem to work for images build with recent versions of the lib build from source. They are still identified as MPO.
Please, ignore the last comment

[cmahnke]

Ok, got the first issue solved, it was an issue with the loading order...

[Yay295]

An MpoImageFile is a JpegImageFile. What is the actual problem you are having?

[radarhere]

If you're working on a pull request to add support for Ultra HDR images, then feel free to revert parts of #8056

[cmahnke]

Thanks, I'll have a look.

But it seems to me that the provided example will overwrite any passed extra, so passing just some binary encoded XMP would get lost there - I need to create my own save_all function.

I don't think this will end up in a pull request, since it's currently quite hacky:

• I did the obvious, which is class UhdrImageFile(MpoImagePlugin.MpoImageFile). But this required me to duplicate seek() and tell() since self.__frame is "private" (two underscores in front) - A clean solution might require changes to MpoImageFile as well. But currently I don't know if my changes are really needed.
• My UhdrImageFile is currently basically just a container, every real operation (like crop, resize and also save) is done using functions not part of the class
• For XMP manipulations I did't do it the pythonic way, which would be a dict based interface, I used str and from there:
• For XMP manipulations I depend on ElementTree's XPATH capabilities - Pillow uses another XML lib

But maybe it's worth to implement a XMP write solution, since this is asked for more often...

[cmahnke]

Ok, XMP itself is a whole other can of worms: One key seems to be #5076, I would like to have a proper XML interface parallel to the existing dict based getxmp() method, just because it's easier to work on data that isn't an abstraction...
I would also ignore the possibility to have the XMP data split over multiple fields in the first try...

The first part, adding XMP to JPEG was far easier then expected:

from PIL._binary import o16be

XMP_MARKER = b"http://ns.adobe.com/xap/1.0/"
XMP_MAX = 65504

def xmp_extra(xml: str):
    if len(xml) > XMP_MAX:
        raise
    length = len(XMP_MARKER) + 2 + len(xml) + 1
    return b"\xFF\xE1" + o16be(length) + XMP_MARKER + b"\0" + xml.encode()

image.encoderinfo['extra'] = xmp_extra(xml)

This gives just a minor warning with exiftool:

exiftool -validate -warning -error -a ../img/xmp.jpeg 
Validate                        : 1 Warning (minor)
Warning                         : [minor] XMP is missing xpacket wrapper

But this also happens with a file generated by the libultrahdr app.

Of course without handling XML larger then 64KB
So easy in fact, that I expect some other pitfalls, otherwise it would be possible with Pillow for years 😉

Next step is to get this into a fork of the MpoImagePlugin._save_all method

[cmahnke]

Well, I'm currently stuck. I have problems figuring out how the MPO writer actually works - I suspect some constants are jus magic numbers? Or better: Field length don't seem to be calculated based on the given input. Trying to pass just XMP as prefix using encoderinfo['extra'] results in a mixup of the MP Part and the ICC Profile...

[cmahnke]

@radarhere : Thanks for #8286. It actually helped me to get a bit further, I hab been blind and didn't noticed that the Image.save()method resets encoderinfo from the image. Anyways, I'm able to add a XMP field to secondary image now, but UltraHDR has another requirement I need to provide:

apply gainmap method expects base image color format to be one of {UHDR_IMG_FMT_24bppYCbCr444, UHDR_IMG_FMT_16bppYCbCr422, UHDR_IMG_FMT_12bppYCbCr420}. Received 2

I suspect to be able to figure it out, just running convert('YCbCr') on the primary image (the one that the append_images array is added to), doesn't seem to do the trick for now.