Multiple-point statistics (MPS) modeling has been the research hotspot of reservoir modeling technology in recent years, and its practicability is restricted by training image. The quality of training image determines the accuracy and reliability of MPS modeling, and is a key factor for the successful application of MPS modeling. This study addresses the characteristics and significance of training image, and systematically introduces the establishment methods of training image creation from the definition, usage and instance, etc., including hand drawn, object-based simulation, 3D seismic information extraction or transformation, prototype- ased model, process-based simulation and 2D image method. Then, our study comprehensively compares the data sources, advantages and disadvantages of various training image creation methods, and discusses the problems caused by MPS modeling relying on training images. Combined with reviewing prior works and MPS modeling practice, the research direction of training image and its establishment methods in the future is pointed out, which provides reference for MPS modeling researchers and users, and provides some thoughts for improving MPS modeling methods.