Abstract: Probabilistic generative models are important methods for knowledge representation. Exact probabilistic inference methods are intractable in these models, and various approximate inferences are required. The variational inferences are important deterministic approximate inference methods with rapid convergence and solid theoretical foundations, and they have become the research hot in probabilistic generative models with the development of big data especially. In this paper, we first present a general variational inference framework for probabilistic generative models, and analyze the parameter learning process of the models based on variational inference. Then, we give the framework of analytic optimization of variational inference for the conditionally conjugated exponential family, and introduce the stochastic variational inference based on the framework, which can scale to big data with the stochastic strategy. Furthermore, we provide the framework of black box variational inferences for the general probability generative models, which train the model parameters of variational distributions based on the stochastic gradients; and analyze the realization of different variational inference algorithms under the framework. Finally, we summarize the structured variational inferences, which improve the inference accuracy by enriching the variational distributions with different strategies. In addition, we discuss the further development trends of variational inference for probabilistic generative models.