Deep neural networks typically demand a substantial volume of labeled training data. Nonetheless, the process of gathering and annotating this training data is frequently labor-intensive and time-consuming. One way to address this issue is through Image Composition. Image composition refers to the process of combining multiple visual elements or components. It may involve techniques like image synthesis or generation, where algorithms generate new images by combining elements from existing ones. This paper aims to overview the recent advancements in Image Composition based on deep networks and relevant applications. It offers an extensive overview of the concepts, principles, and techniques related to Image Composition. Subsequently, it delves into two distinct levels at which image composition can be accomplished: specifically, with 2D elements and 3D objects. Furthermore, the method of achieving composition with point clouds is presented.