Machine learning models often assume abundant training data, but many scenarios lack sufficient samples, especially for certain categories. Zero-shot learning (ZSL) resolves this by expanding the existing knowledge to other areas of knowledge using auxiliary information such as semantic embeddings or attributes. A key challenge is generalizing knowledge through semantic representations while facing domain shift. Current ZSL methods emphasize semantic space but often overlook inter-class discriminability and generalization, leading to perform poorly on fine-grained datasets. To tackle this, we propose a framework that enhances visual-space fidelity and discrimination. We use a discriminator-trained regressor to generate features closer to the real data manifold and increase inter-class distance in semantic space to improve class separability. Our method is compatible with generative ZSL models and improves accuracy. Through comparison on five benchmark test sets, our proposed method has achieved the best results.