In crime investigation, the need to generate accurate facial images from textual descriptions is crucial for identifying suspects and solving cases. Traditional approaches consume more time and accuracy also very low. In recent times, researches using deep learning-based GAN architectures, such as DCGAN and StackGAN, have shown promise in generating realistic images. However, these models often consist of multiple generators, which can introduce ambiguity and complexity in the generation process, particularly when dealing with textual descriptions. So, we proposed a Deep Fusion Generative Adversarial Network architecture to overcome the disadvantages associated with the existing systems. Our model integrates seven new layers termed as UP Blocks to enhance feature extraction and synthesis, while also incorporating a discriminator with matching-sensitive gradient regularization (MS-GR) to improve the discrimination between real and generated images. Through extensive experimentation, we demonstrate the effectiveness of our methodology in producing high-quality facial images that closely align with the provided textual descriptions. For evaluation of the modal, we used Frechet Inception Distance and inception score as metrics and our model achieves the inception score of 1.318 + −0.225 and Frechet Inception Distance score of 30.45, which surpasses the existing models.