A collaborative approach to image generation

Google Research · 9 months ago

Researchers developed PASTA, a reinforcement learning agent that collaborates with users through iterative dialogue to refine text-to-image generation results rather than requiring users to manually adjust prompts. The agent was trained on over 7,000 real human interactions combined with 30,000 simulated interactions, and in human evaluations 85% of raters preferred PASTA's final images over baseline models. The approach enables users to reach their desired image through back-and-forth interaction, with the system learning individual preferences on-the-fly to guide generation toward the user's creative vision.