Semantic image synthesis, involves the transformation of semantic layouts into realistic images, is aimed at comprehending and leveraging given semantic information. Despite recent impressive advancements, challenges persist in terms of fidelity, semantic alignment, and training stability. To enhance the generation quality and semantic alignment in semantic image synthesis, we have reengineered the noise mapping and semantic space embedding, proposing a novel semantic image synthesis model, GAN-Diffusion Relay Model (GDRM), based on GAN and relay diffusion model. Extensive experiments on benchmark datasets validate the effectiveness of our proposed approach, achieving state-of-the-art performance in terms of fidelity (FID) and diversity (LPIPS).