The advancement in computing makes the Generative Adversarial Networks (GANs) models more powerful. There are many fascinating models for generating images from the textual description, but still, the models have not attained the acceptance of the researchers through their reliability and consistency. The work proposes an improved version of Attentional Generative Adversarial Networks (AttnGAN) with Gradient Penalty. It is found in the experiment that the proposed strategy enhanced its performance in terms of Structural Similarity Index (SSIM) and Entity Matching Score (EMS). The visual quality of the output images in the experiments with the CUB-200-2011 dataset is improved and consistent, so that the model achieved 0.75 and 0.95 for SSIM and EMS, respectively. The work highlights the opportunities for cutting-edge research in generative models in the unstructured data generation and interpretation.