The work aims to explore the current status and limitations of artificial intelligence (AI) technology in the field of graphic design and propose a natural language-based image generation method to meet the needs of designers. The GPT-3 model was used to generate textual descriptions of creative ideas and the stable diffusion algorithm was adopted to generate images matching the descriptions. Through practical applications and testing, the effectiveness of the method in graphic design was examined, aiming to discover new perspectives and tools that simplified the design process and enhanced its quality and efficiency. The study adopts an empirical analysis approach, combining experimental results and qualitative analysis to evaluate the level of alignment between generated images and natural language descriptions, image quality, flexibility, and controllability of the method. By utilizing these evaluation criteria, the researchers can determine the method's effectiveness and better fulfill the requirements of designers. Future research directions include enhancing algorithm diversity and creativity, as well as exploring more intelligent graphic design tools to improve the efficiency and quality of designers' work.