The paper introduces a novel generative adversarial network (GAN) called f-CLSWGAN, which synthesizes CNN features conditioned on class-level semantic information to address the zero-shot learning (ZSL) and generalized zero-shot learning (GZSL) tasks. The proposed approach pairs a Wasserstein GAN with a classification loss to generate discriminative CNN features, improving the accuracy of softmax classifiers or multimodal embedding methods. Experimental results on five datasets (CUB, FLO, SUN, AWA, and ImageNet) demonstrate significant accuracy improvements over state-of-the-art methods in both ZSL and GZSL settings. The f-CLSWGAN model is versatile, compatible with different deep CNN architectures and class embeddings, and shows strong generalization capabilities. The paper also highlights the practical application of adversarial training and proposes GZSL as a proxy task to evaluate generative models' performance.The paper introduces a novel generative adversarial network (GAN) called f-CLSWGAN, which synthesizes CNN features conditioned on class-level semantic information to address the zero-shot learning (ZSL) and generalized zero-shot learning (GZSL) tasks. The proposed approach pairs a Wasserstein GAN with a classification loss to generate discriminative CNN features, improving the accuracy of softmax classifiers or multimodal embedding methods. Experimental results on five datasets (CUB, FLO, SUN, AWA, and ImageNet) demonstrate significant accuracy improvements over state-of-the-art methods in both ZSL and GZSL settings. The f-CLSWGAN model is versatile, compatible with different deep CNN architectures and class embeddings, and shows strong generalization capabilities. The paper also highlights the practical application of adversarial training and proposes GZSL as a proxy task to evaluate generative models' performance.