Personalized Image Generation with Reasoning and Reflection
The paper proposes a benchmark for image generators that personalize from a user’s broader history, not just reference images.
It defines two tasks: placing an object into a user-matched scene, and generating a new image on a topic in the user’s visual style. The benchmark draws on e-commerce and social-media settings and evaluates fidelity, quality, distinguishability, semantic alignment, and task utility. The authors also introduce PEARL, a multimodal reasoner paired with a frozen image generator, using a reason-reflect loop. They report a 15% average gain over baselines on personalization metrics. HF Daily Papers' note
It defines two tasks: placing an object into a user-matched scene, and generating a new image on a topic in the user’s visual style. The benchmark draws on e-commerce and social-media settings and evaluates fidelity, quality, distinguishability, semantic alignment, and task utility. The authors also introduce PEARL, a multimodal reasoner paired with a frozen image generator, using a reason-reflect loop. They report a 15% average gain over baselines on personalization metrics. HF Daily Papers' note
score 4