Generalized Zero-Shot Learning (GZSL) seeks to perform classification on both seen and unseen visual instances by exploiting the associations between semantic prototypes and visual features. Generative zero-shot learning synthesizes visual samples of unseen classes using generators conditioned on semantic prototypes. However, predefined semantic prototypes often fail to accurately represent the visual instances of their corresponding classes, which limits the quality of the generated samples. Furthermore, due to the lack of boundary constraints in the generator, some synthesized samples near decision boundaries negatively affect the final classification accuracy. To address the issues of low-quality synthesized samples and boundary sample interference, this paper proposes a Semantic Evolution and Boundary Samples Suppression (SEBS) framework. Specifically, visual samples and semantic prototypes are embedded into a common space, where contrastive learning is employed to transfer visual knowledge to semantic prototypes. The evolved semantic prototypes guide the generator to generate higher-quality visual samples. Additionally, using semantic prototypes as anchors, a gating mechanism is employed to filter out boundary samples that deviate significantly from the anchors. These samples are then suppressed during generation by applying low confidence scores, thereby reducing inter-class confusion. To evaluate the effectiveness of the proposed approach, comprehensive experiments were carried out on three commonly used benchmark datasets. The results show that the proposed method surpasses existing state-of-the-art techniques.