Google researchers announced PASTA, a collaborative approach to image generation that learns what each person actually likes rather than forcing them to master intricate prompts. In PASTA, a user sees four image candidates per turn and simply picks a favorite; the system watches those selections across rounds and rapidly fits a compact model to the individual’s aesthetic preferences. The team trained and evaluated PASTA with a large interaction corpus that includes 7,000 real human sessions and 30,000 simulated sessions, giving the framework broad coverage of styles and goals. In head-to-head comparisons, users reportedly preferred PASTA outputs about 85% of the time over standard baselines, with especially strong gains on abstract or open-ended prompts where “what good looks like” varies by person. Importantly for the community, Google has released both the interaction dataset and the simulation tools so others can build and test systems that personalize to end users. The upshot: as image models get broadly competent, differentiation shifts to user fit. PASTA reframes generation as an interactive loop that quickly homes in on taste, turning prompt roulette into a guided, adaptive workflow that improves with use. Read Google’s research blog for full details: A collaborative approach to image generation.