To address the vastly different styles and insufficiently high fidelity of simulation images simulated by the conventional image-generation platform, a cycle generative adversarial network is proposed. This enables the simulated image in the semi-physical simulation test of an image-guided weapon to closely resemble the actual battlefield environment. The proposed network comprises generators and discriminators, and the method involves an unsupervised deep-learning model embedded in existing image-generation platforms. To ensure real-time performance, locking shared memory technology is used to address image transmission timeouts in semi-physical objects. Test results show that the method can ensure real-time performance and improve confidence in semi-physical simulations.