Nice to see it (especially on other image data than just faces). But typically non-negative matrix decomposition creates less ghostly images, with an easier interpretation.

See e.g. http://www.quantumblah.org/?p=428 or http://scikit-learn.org/stable/auto_examples/decomposition/p... (comparison of PCA and non-negative matrix factorization).

Of course, the non-negative factorization comes at some cost (mostly: much higher computational complexity), but it may be worth trying.