Why is Pillow rotating my image when I save it?

Posted 28 March 2024
Tagged with images, python

Images can have orientation specified in their EXIF metadata, which isn’t preserved when you open and save an image with Pillow.

I was having an issue where an image I’d exported from my Photos Library was being rotated when I opened and saved it:

from PIL import Image  # pip install Pillow==10.1.0

im = Image.open("cross_stitch.jpg")
im.save("cross_stitch_saved.jpg")

The original image had a portrait orientation, but when I saved it the new image had a landscape orientation. How?

After a bit of Googling and head scratching, I found a comment from Andrew Murray in the Pillow repo where he mentioned a possible reason:

I would imagine that the images you are dealing with have an EXIF orientation tag. This means that the image data is saved in one position, and then the image instructs the viewer to rotate it another way.

And indeed, when I opened the failing image in Preview.app and poked around in the metadata, I found a line that suggested this was the case:

Orientation: 6 (Rotated 90°C CCW)

There are two ways to ensure the orientation is preserved when I open and save with Pillow:

Use Pillow to “bake in” the rotation, where you apply the transformation to the pixels:
```
from PIL import Image, ImageOps

im = Image.open("cross_stitch")
im = ImageOps.exif_transpose(im)
im.save("cross_stitch_saved.jpg")
```
This is what I did for my issue, because I didn’t care about any of the EXIF metadata.
Copy the EXIF data into the new photo:
```
from PIL import Image

im = Image.open("cross_stitch.jpg")
im.save("cross_stitch_saved.jpg", exif=im._exif)
```
I’m not sure if this is the right way to copy EXIF in Pillow, because of the underscore on the attribute. It seemed to work in a quick test as I was writing this TIL, but I’d want to look more carefully if I really wanted to preserve the EXIF.

The image dimensions were a clue

I was trying to write a regression test for this issue that checked the image dimensions were preserved upon saving – because if the image is rotated, the dimensions must have changed:

from PIL import Image

im = Image.open("cross_stitch.jpg")
im.save("cross_stitch_saved.jpg")

saved_im = Image.open("cross_stitch_saved.jpg")

assert im.size == saved_im.size

I was surprised when this test passed before I’d fixed the rotation bug!

But the before/after dimensions were the same – 4032 × 3024, which should have been more of a clue. When I opened the image in Preview.app, it was a portrait photo, but those are landscape dimensions! It was a clue that the raw pixel data didn’t match what was being displayed, but I didn’t realise it was significant until after I’d found the underlying issue.