Closed bigcat88 closed 2 months ago
Sorry my bad, I tested it on Pillow 10.2
on Pillow 10.3 there is no error when decoding, just second image looks wrong.
Updated issue to reflect this. Sorry for mess.
Hi, @bigcat88 , 10.3.0 correctly extract the second frame with a different resolution. However, the second frame is not actually a frame, it's a gain map
according to this UltraHDR format.
A JPG file should only have one frame. And people naturally expect the second frame would in the same resolution as the first one. That's the problem we have in https://github.com/comfyanonymous/ComfyUI/issues/3416#issuecomment-2097406107
So, supporting UltraHDR is a good way to go, I guess. #8036
10.3.0 correctly extract the second frame with a different resolution
No it does not, imho.
Pillow should either reject a gain map
and show only primary image, or it should decode second frame correctly, even with lower resolution.
btw, the tomato photo I provided eariler is not very representitive, since the gain map
is all 1.0s (completely white). I didn't understand this was related to the UltraHDR format earlier.
Here is another photo I took that has something in the gain map
:
we can see the gain map
is smaller comparing to the original photo.
btw, the tomato photo I provided eariler is not very representitive, since the
gain map
is all 1.0s (completely white).
please look at my first message. It is not white.
perhaps gain maps need their own type returned from .get_bands() or some way to explicitly exclude them from ImageSequence.Iterator
Are there any images with a small file size that could be added to our test suite and distributed under the Pillow license?
Are there any images with a small file size that could be added to our test suite and distributed under the Pillow license?
I, as the author of the tomato photo and the trees photo, agree to add these to the Pillow test suite and to adhere to its licensing terms.
I don't have smaller ones, since this is already the 'medium resolution'; the other option is 'full resolution.'
Pillow should either reject a gain map and show only primary image, or it should decode second frame correctly, even with lower resolution.
I've created #8056 to treat Ultra HDR images as standard JPEGs, ignoring the gain map. I would consider decoding the second frame correctly to be a matter for #8036
Thanks very much for the test image. I've included it in the PR.
Pillow should either reject a gain map and show only primary image, or it should decode second frame correctly, even with lower resolution.
I've created #8056 to treat Ultra HDR images as standard JPEGs, ignoring the gain map. I would consider decoding the second frame correctly to be a matter for #8036
Im not familiar with the specs for MPO or Ultra HDR, but I am assuming the smaller gain map is embedded in the exif data of the main photo and is being extracted as a seperate file when opening images? Is there any way to reattach the gain map to the exif of the main image so the data is still there later but it is only presented as a single image? I imagine there are future uses where preserving the data is preferred behavior, but presenting them as seperate image is not.
My suggestion will just change the automatic detection of the image format. If a user would like to retain the current behaviour, they could open it directly as an MPO
from PIL import MpoImagePlugin
im = MpoImagePlugin.MpoImageFile('input.jpg')
im.show()
Gotcha, My quick read suggests MPO is MultiPictureObject format, intended for stereographic images, but the Ultra HDR spec hijacks it using the 2nd image as a gain map.
It seems like this being in xmp meta-data separates one of these UltraHDR images from a true MPO stereoscopic file where both images would be wanted:
<rdf:li rdf:parsetype="Resource,
<Container: Item
Item: Semantic="GainMap"
Item: Mime=" image/jpeg"
Item: Length="66171" />
</rdf:li>
It seems like users should not have to be familiar with different implementations of the same file specification to be able to get the proper output when opening a file, so a quick parse of the XMP looking for Item: Semantic="GainMap", should differentiate between a true MPO stereoscopic file, and an UltraHDR file, and dictate if standard opening behavior shows one image, or two.
In #8056, I'm checking for the "Signal of the format", 'hdrgm:Version'.
If you're trying to say that isn't the best way to check if an image is Ultra HDR or not, I'm not following why. Nor does the spec seem to indicate that an Ultra HDR image will be anything other than a single JPEG without a gain map.
Gotcha, I was suggesting that maybe checking both is a good way to guard against improper spec implementation. It seems less likely to me that they forget to label their gain map in the XMP than it is that they specify the hdrgm:Version incorrectly. In any case, they are both in the XMP, and redundancy never hurt anyone.
For the moment, unless you've found a malformed image like that in the wild, I'd rather not double check. When it comes to identifying a format, I don't think it's Pillow's overall style.
If/when https://github.com/python-pillow/Pillow/issues/8036 is resolved, then I expect both the signal and the gain map to be required.
What did you do?
What did you expect to happen?
Correctly decode second image
What actually happened?
What are your OS, Python and Pillow versions?
Test image can be found here: https://github.com/comfyanonymous/ComfyUI/issues/3416#issuecomment-2097406107