HGAN: Editable Visual Generation from Hindi Descriptions

Visual content generation is an active area of research nowadays with considerable work. Still leaving possibilities for enhancement. However, in this area, the domain of the Hindi language has remained unexplored, largely due to the scarcity of linguistic resources. As Hindi is the fourth most spoken language in this world, such work will be a considerable contribution to the research. This paper introduces the HGAN (Hindi Generative Adversarial Network) for creating high-quality visual content from textual descriptions in Hindi and allows the user to change specific image attributes. To ensure diversity, word-level spatial attention and channel-wise attention mechanisms are employed to enable the model to offer control over the attributes of generated content. HGAN shows considerably good results, a 20.3% increase in IS and a 16.6% decrease in reconstruction error on the Hindi dataset than state-of-the-art methods.