Nonnegative Matrix Factorization (NMF) has received considerable attention due to its psychological and physiological interpretation of naturally occurring data whose representation may be parts based in the human brain. However, when labeled and unlabeled images are sampled from different distributions, they may be quantized into different basis vector space and represented in different coding vector space, which may lead to low representation fidelity. In this paper, we investigate how to extend NMF to cross-domain scenario. We accomplish this goal through TNMF - a novel semi-supervised transfer learning approach. Specifically, we aim to minimize the distribution divergence between labeled and unlabeled images, and incorporate this criterion into the objective function of NMF to construct new robust representations. Experiments show that TNMF outperforms state-of-the-art methods on real datasets