Issue arose at the end of bug 53597. Filing a new bug and closing the old one for clarity.

Changing platform from MacOS to ALL since I'm seeing the same thing on Windows XP.

The faint differences that I see are apparently due to the mathematical impossibility of representing some of the color-cube's intensity levels exactly when gamma is 1.0 (which is very different from 1/2.2). Particularly, the sRGB intensity level 0x33 (51) is 7 in gamma=1.0 space and when you convert that back to sRGB the level is 49 instead of 51. The next step, 8, would convert to 52. Similarly, the level 0xcc (204) ends up being 203 when converted to gamma=1.0 when creating the test and back to sRGB by the browser. The only way to "fix" this would be to use greater than 8-bit precision, but we don't want to do that, so I suggest resolving this as WONTFIX or possibly INVALID.

Fair enough - thanks for looking into that Glenn

Differences are no longer faint. They become quite noticeable lately. It's very visible at the last 2 rows here: http://libpng.org/pub/png/png-gammatest.html For example, in the last image in the last row, instead of expected color #CC9900 I see #D1A416