AI image generators still don't know you from a stranger typing the same prompt.
Researchers have published the first unified benchmark for personalized image generation built from a user's actual history, including reviews, posts, captions, and metadata, rather than a handful of curated reference photos. It covers two tasks: placing a product in a scene that matches a shopper's taste, and generating a fresh image on a topic that stays true to someone's established aesthetic, the kind of thing a social app might auto-suggest. Results are graded on five axes, including whether the image looks like the intended target, whether it's visually distinguishable, and how well it matches a given user's history. The team also built PEARL, a system that pairs a multimodal reasoning model with an image generator in a loop that reasons, generates, and reflects before producing a final image.
Most "personalized" image tools today just clone a reference photo's style or face, treating personalization as a curation problem. This work treats it as a context problem, the same bet already reshaping recommendation feeds and ad targeting, and PEARL's 15% average improvement over baselines suggests reasoning-first architectures have real room to beat simple conditioning. For e-commerce and social platforms, that has teeth: a feed that generates product photos or content styled to match your actual browsing and posting history is a layer of personalization most apps don't attempt yet.
It's also a reminder that every caption, review, and post you've left behind is now training data for pictures you never asked an AI to make.