Evolving Knowledge Boundaries in Agentic Visual Generation with SearchGen

Visual generators have demonstrated remarkable capabilities in rendering but struggle with generating content beyond their training data, particularly with evolving user requests that include new characters and events. Researchers have developed SearchGen-20K and SearchGen-Bench, comprising 20,839 prompts across various failure categories and domains, along with a multimodal SearchGen-Corpus-1M to facilitate reproducible research. Current open generators scored poorly on SearchGen-Bench, indicating a significant knowledge gap. The study proposes a teach-then-search co-training framework to improve generator performance by identifying and addressing the evolving knowledge boundaries. The researchers will release their datasets and frameworks to aid in developing more effective, world-knowledge-grounded visual generation systems.