How should Image.fromarray(a, mode='P') be replaced ? there's one occurrence of this in matplotlib

I have some uint8 YCbCr data that I access via imgycbcr = PIL.Image.fromarray(arr, 'YCbCr'). What is the correct equivalent with this deprecation?

Could you put together a simple example to demonstrate what you're doing?

from PIL import Image
import numpy
a = numpy.ones((1, 1), dtype=numpy.uint8)
Image.fromarray(a, "YCbCr")

gives

ValueError: not enough image data

I have a numpy array of shape (h, w, 3) which contains YCbCr data. For instance, a single blue pixel might be

a = np.array([[[30, 255, 107]]], dtype=np.uint8)

I want to convert this to RGB:

ycbcrimg = PIL.Image.fromarray(a, 'YCbCr')
ycbcrimg.convert('RGB').getpixel((0, 0))

yields (0, 1, 255) (our blue pixel).

It looks like I can do

ycbcrimg = PIL.Image.frombuffer('YCbCr', (a.shape[1], a.shape[0]), a.tobytes(), 'raw', 'YCbCr', 0, 1)

but that seems cumbersome compared to the now deprecated PIL.Image.fromarray(a, 'YCbCr'), and I was wondering if there was an alternative preferred idiom for this.

What do you think of this?

from PIL import Image
import numpy as np
a = np.array([[[30, 255, 107]]], dtype=np.uint8)
img = Image.new("YCbCr", (1, 1))
img.putdata([tuple(b[0]) for b in a])
print(img.convert('RGB').getpixel((0, 0)))

I assume that would be inefficient for large images, since we are constructing a tuple for every pixel. I used a 1x1 pixel image as the smallest working example; my actual images are larger and can be numerous.

Ok. If the tuple construction is the only problem, I could put together a PR to allow lists to be used, meaning that

from PIL import Image
import numpy as np
a = np.array([[[30, 255, 107]]], dtype=np.uint8)
img = Image.new("YCbCr", (1, 1))
img.putdata(a.tolist()[0])
print(img.convert('RGB').getpixel((0, 0)))

would work. How does that sound?

We really go out of our way to tell people not to do bulk pixel manipulation in python, so deprecating something that does the right thing in 0 time in C and then suggesting that level of pixel pushing in Python seems wrong.

When this was proposed, there was the one legitimate use case mentioned (P), this is definitely another. Specifying that the uint8 pixels represent YCbCr instead of RGB is the sort of cast that is safe and useful for importing -- it's essentially the same operation as specifying a colorspace -- rather than specifying a bit layout which is what trips people up.

I can see two options here:

1. We allow the mode parameter, but only for modes that are compatible with the dtype. This allows for the matrix of RGB(XxA)/Lab/YCbCr to be specified properly. We explicitly error on things like bool->1 and float->int modes. This makes it explicitly cast only, which should remove the footgun aspect of this parameter.

2. We provide an interface to change the im->mode between bit compatible modes, so that one can take any RGB image and interpret it as YCbCr.

Ok, I've created #9063 for the first option.

Pinging @guillaume-rochette-oxb just to make sure he is aware of this potential step in the other direction.

Thank you.

Also ran into this deprecation with pypdfium2.
Problem is, I think this would lose the distinction between RGBA and RGBX, won't it? The former cannot be saved as JPEG, but the latter can.
AFAICS, this case is not included in the revert list in #9063 (comment) yet.

Ok, I've added RGBX to the list of RGBA alternatives in the PR.

Pillow 12.0.0 has now been released with #9063