Imaginique Expressions: Tailoring Personalized Short-Text to Image Generation Through Aesthetic Assessment and Human Insights

Text-to-image tasks have gained significant progress, where vivid images are generated from detailed text descriptions. However, existing studies overlook scenarios where the provided text is sparse, neglect the investigation of human preferences, and fail to acknowledge the diversity of aesthetic opinions. To address those issues, we develop a methodology called personalized short-text-to-image generation through aesthetic assessment and human insights. We develop a Personality Encoder (PE) to extract personal information and establish the Big-Five personality traits-based Image Aesthetic Assessment model (BFIAA) for predicting specific human aesthetic preferences. Utilizing the BFIAA model, we fine-tune the Stable Diffusion model to align it more closely with human preferences. Our experiments demonstrate that our BFIAA model can truly reflect human aesthetic preference and the adapted generation model can generate personalized images more preferred by humans.