The scarcity of SAR images and their time-consuming and laborious labeling is a pressing issue in the development of SAR detection algorithms. In addition, the existing generation algorithms do not consider the complex context of real SAR images, resulting in low-quality of the results. To solve these difficulties, we introduce a SAR image generation model consisting of a scene fusion module, generator and discriminator called SF-SARnet. The model accepts target location information as input and generates a mask map, which is then fed to the network for generation. Our main contributions include the following: (1) We design a generator based on the traditional encoder-decoder structure, on top of which we add jump connections at each odd layer. (2) To increase the quality of the fake generation results, we design a loss function that incorporates perceptual loss. (3) To verify the model's performance, comparative verification was performed on the HRSID dataset. The results indicate that our model outperforms the others. Moreover, the generated images can be conveniently used for other related tasks.