Public service announcements (PSAs) play an important role in guiding correct social values and in the construction of spiritual civilization, but they encounter obstacles in the process of social marketing transformation. From the perspective of network new media, this paper interprets the core elements of PSA design in a correlative way, proposes a three-dimensional design path combining content, communication and interaction, and combines generative adversarial network and VGG16 network to construct an automatic generation model of PSA layout pictures to realize the dynamic optimization of PSA communication effect. At the same time, a collaborative filtering recommendation algorithm based on cosine similarity is used to realize the personalized recommendation of PSAs, in order to achieve efficient and accurate matching between the audience and the advertisement content.The optimal number of iterations of the GAN model is 70 epochs, and at this time, the total loss value of the model is 0.831. The mean values of the highest scores of rationality, aesthetics and typographical neatness of the multi-group layout images generated by the model in this paper are 8.49, 7.98, and 8.67, respectively, and the overall quality of the generated layouts is excellent. Finally, seven main factors for evaluating the perceived value of PSA communication are extracted, namely: social factor, functional factor, emotional factor, style factor, purpose factor, advertising performance factor, and instant factor. This paper provides a useful reference for the overall planning and sustainable development of PSAs and public welfare programs in the context of online new media.