This paper studies the creative generation and evaluation model development of artificial intelligence (AI) in graphic design, aiming at improving the efficiency and quality of graphic design through advanced AI technology. With the intensification of market competition and the diversification of customer needs, graphic designers are facing great challenges, and it is the key to quickly generate high-quality creative designs. This article focuses on Generative Adversarial Network (GAN) technology and designs a creative generation model framework based on GAN. Through data preprocessing, generator network, discriminator network, and adversarial training process, it achieves the generation of diverse creative design images from random noise. In order to increase the diversity and controllability of creativity, the model introduces conditional variables and style transfer algorithm, which enables the generator to generate corresponding creative designs according to different design themes and styles. In the development of evaluation model, this paper constructs an evaluation index system including three dimensions: creativity, aesthetics and practicality, and uses machine learning algorithm to train the evaluation model to quantitatively evaluate the generated creative design. The evaluation model can achieve a comprehensive and objective evaluation of creative design by extracting the characteristics of design samples and learning the relationship between characteristics and evaluation. The experimental results show that the creative generation and evaluation model proposed in this paper is excellent in generating diversified and high-quality creative designs, and the evaluation model has achieved high scores in three dimensions: creativity, aesthetics and practicality. The research in this paper not only provides graphic designers with a new method to quickly generate diversified creative designs, but also reduces the burden of manual review through automatic evaluation model, and improves the design efficiency and accuracy.