In generative zero-shot learning, the visual distribution mismatch between synthetic and real samples is a challenge due to the lack of effective constraints for unseen class visual features. To address this, we propose the Contrastive Prototype Network (CPNet). CPNet uses prototype learning to determine the feature vector center for each category (the prototype) and classifies based on the similarity between test data and prototypes. Concurrently, contrastive learning ensures that representations of the same category are more compact, while those of different categories are more dispersed. Experimental results on three standard datasets confirm the effectiveness of our approach.