What's this about?
Image generators confidently fabricate what they don't know — new characters, trending entities, anything after their training cutoff. This paper makes generation *agentic*: the system detects when a request exceeds its knowledge, searches the web for what the thing actually looks like, and grounds generation in what it finds. HF 32↑.
What exactly is the "knowledge boundary"?
Every model is frozen at training time; the world isn't. The knowledge boundary is the edge between what the model learned and everything newer or rarer. Text models handle it with retrieval (RAG); this brings the same move to image generation — evolving the boundary at inference time instead of hallucinating past it.
Why is it worth noting?
"Agentic generation" — generators that research before they draw — is a natural next step now that agent scaffolding is cheap, and it targets the most embarrassing failure mode of image models.