<p>Text-to-image synthesis is creating a sensible image from a given text description, and creating new pictures from some image description is known as image synthesis. The study of text-to-face generation is severely constrained due to a lack of datasets. Up until now, generative adversarial networks have been used for the majority of text-to-face generation efforts. The pre-trained text encoder has been employed in these networks to extract the semantic components of the input sentence. These semantic properties were later used to train the picture decoder. The ability to create high-resolution images with various objects and the development of appropriate and trustworthy evaluation criteria that correlate with human judgement are two difficulties that the text-to-image synthesis sector still must overcome. To overcome this challenge, we propose a novel enhanced Multinomial Statistical Predictive Mixture Model (MSPMM) that creates realistic and real images. The suggested technique simultaneously trained the text encoder and image decoder to produce more accurate and effective outcomes. The dataset has also been labelled using the classes that we have established. Various sorts of trials were conducted, and the results showed that our suggested MSPMM excelled at them by creating images of excellent quality regardless of the input language. Additionally, the visual results have supported our experiments by generating facial images that correspond to the given query. The results show that the proposed novel MSPMM can create photorealistic face images from attribute labels, and these images can significantly improve attribute recognition model performance by serving as augmented training models. Compared to existing methods such as Stack GAN, GAN, MMC-GAN, our proposed method outperforms them well. The suggested MSPM model obtained an FSD score of 1.018 and an FSD score of 40.06, which are lower than those obtained by other existing methods (Stack GAN, GAN, MMC-GAN).</p>