I had to revert the rebaseline bot's submission recently.  The revert is here:
https://codereview.chromium.org/2182653004/

At this cycle on the WebKit Win10 bot,
https://build.chromium.org/p/chromium.webkit/builders/WebKit%20Win10/builds/12825
the rebaseline bot submitted a change to revise many tests, including the
virtual/gpu-rasterization/fast/images/color-profile-reflection.html
and the
virtual/gpu-rasterization/fast/images/png-with-color-profile.html
webkit test.

This submission did one, maybe two, things wrong.

(1)
The new image it checked is as expected for the
virtual/gpu-rasterization/fast/images/color-profile-reflection.html
test is blank.
I spotted this using the comparison here of expected versus actual.
https://storage.googleapis.com/chromium-layout-test-archives/WebKit_Win10/12825/layout-test-results/results.html
I assuming from the code review that it's this file (though I'm surprised to see Win7 in the name when it's the Win10 that's running the test)
https://codereview.chromium.org/2182653004/diff/1/third_party/WebKit/LayoutTests/platform/win7/virtual/gpu-rasterization/fast/images/color-profile-reflection-expected.png

I reverted the auto-rebaseline submission.

When the revert when through here on this cycle,
https://build.chromium.org/p/chromium.webkit/builders/WebKit%20Win10/builds/12831
which returned the file
third_party/WebKit/LayoutTests/platform/win7/virtual/gpu-rasterization/fast/images/color-profile-reflection-expected.png
to its previous state, the failing webkit test turned green again.

(2)
The new image it checked in for
virtual/gpu-rasterization/fast/images/png-with-color-profile.html
differs in subtle (aliasing) ways from the expected image.

See the comparison here of expected versus actual.
https://storage.googleapis.com/chromium-layout-test-archives/WebKit_Win10/12825/layout-test-results/results.html

This failed consistently until this cycle.
https://build.chromium.org/p/chromium.webkit/builders/WebKit%20Win10/builds/12830

It's not clear to me what fixed it.  There was a new rebaseline in this cycle, but that rebaseline didn't touch the file in question.  There was another change that cycle that could be relevant (#7920).  Regardless, it's clear that the rebaseline bot caused this test to start failing.  (There's nothing else relevant in the first bad cycle that seems relevant.  If the latter change fixed the issue, then perhaps the rebaseline bot computed the expected image with the latter change but somehow checked in in beforehand.  Just hypothesizing.)