Generating Apparel Images by Using Stable Diffusion with Prompt Recommendation

抄録

Fashion is a profound expression of personal identity, style, and culture in daily life. Making oneself look attractive, which often equates to being fashionable or stylish, is always a priority for many people. In this paper, we proposed an apparel image generator based on a pre-trained Stable Diffusion model to assist users in discovering or designing their desired outfits. Because the non-intuitive causality between input text and output image in current generative AI models, many users are often frustrated at the input stage. To enhance users’ prompt selection, we provide a recommend keyword list for users by analyzing the relationships among (1) user’s original prompts, (2) the reverse engineered prompts with the image scores as weighting values, (3) and the fashion related keywords in our database. The above approach is repeated iteratively until users obtain an image result which is significant enough for them. Finally, a feedback survey is also performed. The experimental results show that our system can significantly improve the user experiences and provide more satisfactory cloth images.