A collaborative approach to image generation
Google Research
Google built an AI agent called PASTA that keeps refining images with you turn by turn instead of one-shot prompting. Raters preferred its results over standard text-to-image models 85% of the time.
Based on reporting by Google Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anyone who's fought with a text-to-image model knows the drill: type a prompt, get something close-but-off, rewrite the prompt, repeat. Google Research thinks that loop is the wrong interaction model entirely, and its new system, PASTA, tries to replace trial-and-error typing with an actual back-and-forth.
PASTA stands for Preference Adaptive and Sequential Text-to-image Agent, and it works by generating four image options per turn rather than one. You pick whichever is closest to what you want, and that choice becomes feedback the agent uses to steer its next batch of suggestions. Under the hood it pairs a Gemini Flash model for generating prompt expansions with Stable Diffusion XL for rendering images, and a reinforcement learning agent — trained with implicit Q-learning — decides which four expansions are worth showing you at each step.
The harder problem was getting enough training data without hoovering up real user behavior at scale. Google's team collected just over 7,000 genuine rater interactions, then used that seed data to build a user simulator: a model that predicts both how much someone will like a given image set and which one they'd pick from it. That simulator clustered raters into latent "user types" — people who gravitate toward animals, scenic shots, abstract compositions, and so on — and used those patterns to generate more than 30,000 synthetic interaction sequences.
The testing results are the interesting part. An agent trained purely on synthetic data underperformed a plain baseline. One trained only on the small real dataset improved things but still didn't beat the baseline outright. Only the version blending real and simulated data actually won, and when researchers had raters directly compare its final images against the baseline's, 85% picked PASTA's output. The gap widened most on vague prompts — feeding it something as abstract as "an image of love" produced everything from tender portraits to geometric abstraction, depending on which user type the agent inferred it was talking to.
Google has open-sourced both the rater dataset and the simulated interaction data, framing this less as a finished product and more as a template other generative tasks could borrow — the core idea being that a well-built user simulator can substitute for data you can't ethically or practically collect at scale.
My take — AI-written commentary, not fact-checked reporting
The clever bit here isn't the RL agent, it's admitting that real preference data is scarce and privacy-sensitive, then building a simulator instead of pretending you can scrape your way around that problem. That's a more honest approach to personalization than most 'personalized AI' pitches I see, and I'd rather see labs open-source interaction datasets like this than another benchmark nobody trusts.
Read more about this at: Google Research
