HTGAN: An architecture for Hindi Text based Image Synthesis
Anil Singh Parihar, Aditya Kaushik, Aditya Vikram Choudhary, Amit Kumar Singh · 2021
Synthesis of high-quality images that are semantically consistent with their text descriptions has emerged as an essential and onerous problem in the areas of CV and NLP. Various significant steps have been taken in the task of generation of images using English textual descriptions, with Multimodal GANs being at the forefront of all these efforts. In this work we attempt to extend the existing English text to image generation (T2I) techniques to the novel task of Hindi T2I using language translation models. To achieve this, the input Hindi sentences were translated to English using a transformer based Neural Machine Translation module the output of which was then fed to the Generative Adversarial Networks based Image Generation Module. The outcomes of this approach have been evaluated using well established metrics such as Inception Score and BLEU Score which yield scores that indicate the generation of lifelike images which semantically align with the input text descriptions.