JPEG saved with keep_rgb=True and dpi reopens with incorrect colors #10344

What did you do?

Saved the same RGB image to JPEG with keep_rgb=True, once with DPI and once without it, then reopened both files:

from io import BytesIO

from PIL import Image

image = Image.new("RGB", (16, 16), (255, 0, 0))
for dpi in (None, (72, 72)):
    out = BytesIO()
    options = {"keep_rgb": True, "quality": 100, "subsampling": 0}
    if dpi is not None:
        options["dpi"] = dpi
    image.save(out, "JPEG", **options)
    out.seek(0)
    with Image.open(out) as reopened:
        print(dpi, reopened.getpixel((0, 0)))

What did you expect to happen?

Both images should reopen approximately red. Setting DPI should not visibly change the colors.

What actually happened?

On both installations below, the output was:

None (255, 0, 0)
(72, 72) (76, 255, 28)

The DPI version contains both a JFIF APP0 marker and an Adobe APP14 marker with transform 0; the version without DPI contains only the Adobe marker. As a diagnostic, removing the JFIF marker from the generated bytes made Pillow reopen the pixel as (255, 0, 0), but also removed the DPI. I don't know whether the right fix is in the writer, the reader, or both.

FFmpeg 9.0.2 decodes both generated JPEGs as (255, 0, 0) on the same machine. This suggests a difference in decoding the DPI version; I have not isolated the responsible layer.

The generated file has component IDs R, G and B. As a diagnostic, forcing the low-level JPEG decoder input mode to RGB also restores the expected red in both Pillow versions.

Looking at libjpeg-turbo 3.2.0, its default color-space selection gives the JFIF marker precedence over Adobe APP14, selecting YCbCr when both are present. Pillow enables JFIF when both DPI values are positive.

#7679 notes that setting DPI forces JFIF even for non-JFIF color spaces. The JPEG encoder source is identical between the 13.0.0 tag and current main, although I haven't built main to reproduce this there.

What are your OS, Python and Pillow versions?

• Windows 11, Python 3.15.2, Pillow 13.0.0, libjpeg-turbo 3.2.0.
• Also reproduced with Python 3.12.14, Pillow 12.3.0, libjpeg-turbo 3.1.4.1.

Reproducer and investigation assisted by Codex.