Xiaoyun,
> My question is that if we can only add one regressor of variation for
> the speciafied face category when I focus on the face category, and
> the same for other face categories.
If I understand you correctly, you are concerned about the possibility that the variation relative to each face category might be highly correlated with each other. That is a legitimate concern because it might have effect about the numerical accuracy of the program. On the other hand, if you do want to regress out those variations for all categories, you may go ahead and modulate all categories, and let the program handle the numerical aspects. If the program doesn't whine, you should be fine.
Gang