Personalizing Text-to-Image Generation to Individual Taste

A research team from KU Leuven and the University of Tübingen has published a paper introducing PAMELA, a novel framework designed to tackle a core limitation of modern text-to-image (T2I) models like Flux 2 and Nano Banana. While these models produce high-fidelity images, they are optimized for broad, average human appeal, ignoring the inherent subjectivity of personal taste. PAMELA addresses this by creating a predictive model trained on a rich, new dataset of 70,000 personalized aesthetic ratings across 5,000 diverse AI-generated images, each evaluated by 15 unique users.

The core innovation is a personalized reward model that learns individual user preferences across domains like art, design, and photography. The researchers demonstrate that this model can predict what a specific person will like more accurately than current state-of-the-art methods can predict general population preferences. This capability enables practical applications, such as using simple prompt optimization techniques to steer the output of T2I generators toward visuals that align with a user's unique aesthetic judgment, rather than a generic average.

By releasing both the dataset and the model, the team aims to establish a new standard for research into personalized AI alignment and subjective quality assessment. This work highlights that improving AI's understanding of human preferences requires a shift from monolithic 'goodness' metrics to nuanced, individualized models, fundamentally changing how we interact with and control generative visual AI.