I've created sambvfx#1 to add some more test images - a PNG to compare the image against, and a TYPELESS variation of your test image.

To be certain, this is what your image is supposed to look like, yes?

Thanks @radarhere, I've merged your changes.

One odd thing to note is that the original PNG I used was simply the ati1.png from the existing test images. I converted that to the bc4_unorm.dds using nvidia texture tool exporter like this:

nvtt_export.exe ati1.png --format bc4 --output bc4_unorm.dds

Tweaking the tests to compare against the ati1.png instead of the new one you added fails... There's a chance I'm doing something wrong, but I find that a little unexpected.

The compression is lossy format, so it is not unexpected that converting to it created a slightly different file.

Testing, I find that assert_image_similar_tofile(im, TEST_FILE_ATI1.replace(".dds", ".png"), 0.4) passes, but this assert_image_equal_tofile is a stronger test, so the current version is fine to my way of thinking.

With #6486, this change is now mentioned in the release notes - https://pillow.readthedocs.io/en/latest/releasenotes/10.2.0.html#other-changes