Stylized dialogue generation is capable of producing highly creative and engaging responses, making it an indispensable feature of intelligent dialogue systems. However, a major challenge to this task is the paucity of supervised data, resulting in suboptimal performance. Although some unsupervised methods have emerged, they tend to handle only a limited range of dialogue styles simultaneously. Retraining becomes necessary when new dialogue styles are introduced, leading to increased training overhead and model redundancy. Furthermore, the large language model shows exciting performance, it still falls short on some specific tasks. To address the data limitations and training overhead, we propose a Multi-Stylized Adapter Dialogue Generation (MultiSADG) model, which generates multiple stylized adapters by using representations from different stylized corpora. Specifically, we generate style-specific adapters for modeling both contexts and stylized responses. These style-specific adapters are generated by a hypernetwork trained on multiple stylized corpora. In addition, for unseen stylized texts, MultiSADG uses the stylized corpus to generate style-specific adapters through the hypernetwork to deal with the zero-shot scenario. MultiSADG is evaluated on five stylized dialogue data, and the experimental results show its satisfactory performance in both automatic and manual evaluation.