Just wild speculation as I haven't played with that feature:  some textures take longer to calculate.  If you want to test that hypothesis, see if you can find some correlation with the ones that were nearly instant, the ones that took a few seconds, and the ones that took a minute or two.  Since it's a simple V flip, complexity of the image probably isn't a factor.  Size, OTOH might well be an issue, especially as it probably has to buffer its translation efforts in a duplicated copy.  If size doesn't seem to be an issue, check to see if there are alpha channels in some and not in others.

That's the best I can offer.  Hopefully someone who know what they're talking about will come along and help.