It's pretty weird, we are generating a diff image that says there are differences, but looking at the actual and reference png images they look exactly the same. It could be a problem with the way we create the results for reftests. fast/sub-pixel/sub-pixel-composited-layers.html http/tests/misc/slow-loading-animated-image.html
This is not the case of tests failing because of one pixel, in this case the diff image shows significant differences