Generalized zero-shot learning (GZSL) is a challenging class of vision and knowledge transfer problems in which both seen and unseen classes appear during testing. Most existing GZSL methods achieve knowledge transfer based on the original features of samples that inevitably contain information irrelevant to recognition, resulting in negative influence for the performance. In this paper, we propose a novel contrastive disentanglement learning framework for the GZSL task (SDCE-GZSL), where the original and generated visual features are factorized into semantic-consistent and semantic-unrelated representations via a novel mutual information (MI)-based constraint. In addition, we propose a contrastive learning framework that leverages class-level and instance-level supervision to further facilitate disentanglement. Extensive experiments show that our approach achieves significant improvements over the state-of-the-art approaches.