<p>Most of the existing studies on the intellectualization of traditional culture focus on fields such as image recognition and style transfer. The research on intelligent generation of traditional cultural art design (CAD) texts is scattered and lacks systematicness. To this end, this study conducts research on intelligent text generation in traditional CAD based on the attention mechanism. First, the research status and limitations of the attention mechanism, intelligent text generation, and digitalization of traditional culture are sorted out. Then, a multi-source dataset of traditional art design feature-text pairs is constructed. Next, an improved Transformer model is designed. Based on the standard encoder-decoder framework, a cultural semantic attention layer, a multi-scale temporal convolutional network style constraint module, and an output verification layer are added. Meanwhile, the dynamic weight calculation of the cultural semantic attention layer and the multi-scale feature fusion of the style module are optimized. The results show that the Bilingual Evaluation Understudy-3 (BLEU-3) score of the proposed Attention-based CAD (Attn-CAD) model reaches 0.72. The cultural terminology accuracy is 92.5%, and the style matching degree is 89.3%, with the best performance in the ceramic field. Ablation experiments verify the necessity of the core modules, and the decline rates of indicators in cross-domain tests are all less than 5%, which proves the model’s universality. In terms of methodology, this study solves the core problem of weak semantic association of the general model in traditional cultural scenarios and insufficient terminology accuracy of general models in traditional cultural scenarios; it provides technical support for the living inheritance and digital innovation of traditional culture.</p>