HDR (10-bit) frames cannot be converted to images correctly since av==16.0.0.

[nagadomi]

Since 16.0.0, frame data cannot be converted to images correctly with code like the following.

import av
import numpy as np
from PIL import Image


input_file = "tmp/test_videos/Hybrid HDR10 with Dolby Vision Sample.mkv" # from https://kodi.wiki/view/Samples
output_file = f"{av.__version__}_out.png"
skip_frame = 10


with av.open(input_file, mode="r") as container:
    for packet in container.demux(video=0):
        for frame in packet.decode():
            skip_frame -= 1
            if skip_frame <= 0:
                frame = frame.to_ndarray(format="rgb48le")
                frame_float = frame / 65535.0
                # Save as 8-bit for debugging.
                frame_uint8 = np.clip(frame_float * 255.0, 0, 255).astype(np.uint8)
                image = Image.fromarray(frame_uint8)
                image.save(output_file)
                print("save", output_file)
                break
        if skip_frame <= 0:
            break

15.1.0

16.0.0 and 16.1.0

If the issue is related to the reformat or to_ndarray, it may already be fixed in the latest code, but I have not tested it.

[lgeiger]

I can reproduce this on main as well:

[nagadomi]

@lgeiger Thanks for checking. I noticed that the arguments in the code above were incorrect and fixed them, but the result was the same.

[nagadomi]

I found that it works correctly when gbrp16le is specified.
The issue appears to be related to rgb48le handling.
For some reason the result is slightly different (PSNR: 53), but using gbrp16le seems to avoid this issue for me.

EDIT: gbrp16le appears to be significantly slower than rgb48le.