Zero-shot learning (ZSL) addresses the unseen class recognition problem by\nleveraging semantic information to transfer knowledge from seen classes to\nunseen classes. Generative models synthesize the unseen visual features and\nconvert ZSL into a classical supervised learning problem. These generative\nmodels are trained using the seen classes and are expected to implicitly\ntransfer the knowledge from seen to unseen classes. However, their performance\nis stymied by overfitting, which leads to substandard performance on\nGeneralized Zero-Shot learning (GZSL). To address this concern, we propose the\nnovel LsrGAN, a generative model that Leverages the Semantic Relationship\nbetween seen and unseen categories and explicitly performs knowledge transfer\nby incorporating a novel Semantic Regularized Loss (SR-Loss). The SR-loss\nguides the LsrGAN to generate visual features that mirror the semantic\nrelationships between seen and unseen classes. Experiments on seven benchmark\ndatasets, including the challenging Wikipedia text-based CUB and NABirds\nsplits, and Attribute-based AWA, CUB, and SUN, demonstrates the superiority of\nthe LsrGAN compared to previous state-of-the-art approaches under both ZSL and\nGZSL. Code is available at https: // github. com/ Maunil/ LsrGAN\n