[saintparish4]

Regression from #10210

#10210 started ignoring a PhotometricInterpretation tag with more than one value, by not setting it in _setitem(). However, the raw data is still kept in _tagdata, so the tag is reported by in and when iterating, but KeyError: 262 is raised when it is read.

This means that an image with such a tag can be opened and loaded, but reading its EXIF data or saving it as a TIFF fails.

import io
from PIL import Image

buf = io.BytesIO()
Image.new("L", (4, 4)).save(buf, "TIFF")
b = bytearray(buf.getvalue())
offset = int.from_bytes(b[4:8], "little")
for i in range(int.from_bytes(b[offset : offset + 2], "little")):
    entry = offset + 2 + 12 * i
    if int.from_bytes(b[entry : entry + 2], "little") == 262:
        # Change PhotometricInterpretation to have two values
        b[entry + 4 : entry + 12] = b"\x02\x00\x00\x00\x01\x00\x02\x00"

with Image.open(io.BytesIO(b)) as im:
    im.load()
    im.save(io.BytesIO(), "TIFF")  # KeyError: 262

Changes proposed in this pull request:

• Skip a PhotometricInterpretation tag without exactly one value when loading the IFD, so that its data is not stored. This is only done for the main IFD, as tag 262 does not have a fixed length in the Exif, GPS and Interop IFDs.
• Added a test.

If a malformed PhotometricInterpretation tag is followed by a valid duplicate, the valid one is now used, whereas after #10210 neither was. Let me know if you would prefer to keep ignoring both.

[radarhere]

like libtiff

I'm intrigued. Could you expand on how you know this, link to source code, or something?

[saintparish4]

It's from libtiff's directory reading. PhotometricInterpretation is defined as a single SHORT (tif_dirinfo.c#L104) and has no special case in TIFFReadDirectory, so it's read by TIFFFetchNormalTag(tif, dp, TRUE) (tif_dirread.c#L4896-4897). That calls TIFFReadDirEntryShort, which returns TIFFReadDirEntryErrCount if the count isn't 1 (#L485-489), so TIFFSetField is never called (#L6517-6529). Because recover is set, it only warns, Incorrect count for "PhotometricInterpretation"; tag ignored (#L4063-4067).

Looking into this also answered my question about duplicates. libtiff marks every later duplicate of a tag as ignored before reading any values (#L4389-4408), so a malformed PhotometricInterpretation followed by a valid one ends up with neither. I've pushed a commit so that Pillow does the same, with a test.

This only covers PhotometricInterpretation, though. When Pillow skips the first copy of another tag (unsupported type, empty data, etc.), a later duplicate is still used, unlike libtiff. I'm happy to make that general, here or in a follow-up, if you'd prefer.