Addressing Bias in Text-to-Image Generation: A Review of Mitigation Methods

Although text-to-image algorithms are quite good at converting text to images, they may use biased training data. This study looks into several definitions of bias and methods for calculating and mitigating it in models that convert text to images. We evaluate previous research on gender, skin tone, and geo-cultural bias in this survey. It shows the positive and negative aspects of different methods for evaluation and mitigation of bias. We discovered that: (1) To prevent misunderstandings when talking about "bias," it is crucial to make sure that definitions of bias are based on societal harms and to recognize the similarities and differences across suggested definitions; (2) Different sets of handmade prompts are commonly used, but there are no common assessment frameworks or metrics, and there are no unifying criteria for evaluation; and (3) It is necessary to conduct an extensive study to find bias mitigation methods to provide fair images because current mitigation techniques are unable to completely address biases. We identify future research directions that support precise definitions, evaluations, and mitigation of bias based on the present constraints. By highlighting the significance of researching biases in T2I systems, we want to inspire more research into the comprehensive understanding and management of biases, with the ultimate goal of creating equitable and reliable T2I technologies for all.