Open Access

Maria-Teresa De Rosa Palmini, Laura Juliane Wagner

• Text-to-image generative models offer an innovative method for creating visual content, exploiting the limitless potential of text-based inputs. However, the reliance on text prompts can lead to a labor-intensive process of experimentation for users aiming to achieve high-quality results. This has led to the development of a specialized prompt language with specific, descriptive keywords that users can exploit so as to achieve the best possible visual outcomes. Insight into prompting strategies can be obtained by analyzing the media generated and shared on text-to-image online platforms. Utilizing natural language processing (NLP) and visualization techniques, a detailed analysis of the prompts that led to the creation of images with exceptional popularity (by like count) was performed. The present study is focused on identifying the predominant topics and language patterns that contribute to the creation of images that receive high community ratings. Our analysis reveals a strong focus on surface aesthetics in prompts, which often emphasize conventional beauty and popular visual themes and a gravitation towards erotic and pornographic content. Positive prompts typically involve descriptions of female bodily features, blending elements of fantasy and realism. Negative prompts consistently counteract what seems to be perceived as visual imperfections, often describing body horror, marked by distorted human features as well as technical imperfections related to digital imagery in particular.