Enhanced Text-to-Image Synthesis with Self-Supervision

The task of Text-to-Image synthesis is a difficult challenge, especially when dealing with low-data regimes, where the number training samples limited. In order to address this Self-Supervision Generative Adversarial Networks (SS-TiGAN) has been proposed. method employs bi-level architecture, which allows for use self-supervision increase by generating rotation variants. This, in turn, maximizes diversity model representation and enables exploration high-level object information more detailed image construction. addition self-supervision, SS-TiGAN also investigates various techniques stability issues that arise Networks. By implementing these techniques, proposed achieved new state-of-the-art performance on two benchmark datasets, Oxford-102 CUB. These results demonstrate effectiveness synthesizing high-quality, realistic images from text descriptions under regimes.