Alpha channel lost when convert between pillow image and videoframe

[hermeschen1116]

Hello, I recently found alpha channel will lost when convert between pillow image and videoframe.
I've already checked the source code, and I found the alpha channel will be ignored during conversion.
Is this intentional or a bug? Will it be implemented later?

[abdullayevziyodulla]

The current Pillow convenience methods use RGB: from_image() constructs an rgb24 frame, and to_image() produces an RGB Pillow image. Consequently, those methods discard alpha even if your original Pillow image is RGBA. See the implementation.

You can preserve it through the ndarray APIs instead:

import av
import numpy as np
from PIL import Image

rgba = np.asarray(image.convert("RGBA"))
frame = av.VideoFrame.from_ndarray(rgba, format="rgba")
restored = Image.fromarray(frame.to_ndarray(format="rgba"))

I checked this with PyAV 19.0.1 and a synthetic RGBA image: the round trip preserved (10, 20, 30, 47), including the alpha value. Calling VideoFrame.from_image on the same image produced rgb24.

This addresses the image/frame conversion. If you later encode the frame, your codec, pixel format and container also need alpha support; converting it to a format such as yuv420p will lose alpha. I can't speak for whether the maintainers intend to extend the convenience methods, but the explicit rgba route avoids that limitation today.