Real valued Boltzmann machines

I've been reading about RBMs lately and one thing I realize I'd that almost all the examples assume binary visible units. The one exception was the Gaussian-binary RBM which can cope with real valued input. But it looked too complicated and much style top train. How do people use RBMs on such a wide variety of real valued data sets? Audio, natural images etc. Do they all use Gaussian-binary RBMs or am I missing something simple?