Abstract:The Generative Adversarial Network (GAN) can generate images with significant complexity and authenticity, but it is usually constructed to sample from a single potential source, thus ignoring spatial interactions between multiple entities that may exist in the scene. To capture the complex interactions between different objects, including their relative scaling, spatial layout, occlusion or view conversion, in this paper we propose an image condition-based generative adversarial network by using decomposition-synthesis procedures. The model can generate realistic composite images from their joint distributions based on their texture and shape of the input object. Through use of Shapenet data set, we experiment on 51 300 image models out of the respective 55 common objects in the 2D and 3D images. Compared with the traditional SLP and cGAN, the image quality in our algorithm can be improved by 4%.