Dreams have been responsible for some major creative and scientific discoveries in the course of human history. Recording dreams in the form of images is an interesting and meaningful thing. Our task is to generate images based on the description of dreams. Recently, there has been exciting progress in generating images from descriptions of birds and flowers, but the dream scene is more fantasy than the real scene. The challenge to reproduce complex sentences with many objects and relationships remain. To truthfully reappear the dream scene, we process sentences into scene graphs that are a powerful structured representation for both images and language; then using a graph convolution network to obtain layout information, combining the layout information and a single feedforward network to generate the image. Subsequently, we apply Cycle-Consistent Adversarial Net (CycleGAN) to change the image into different styles according to the mood of users when dreaming. According to the experimental results, our method can generate complex and diverse dreams.