Standard text-to-image models are limited to what they saw in training. Agentic image generation tries to fix that by pairing a generator with a web search agent that retrieves reference images. But real-world requests often concern events or people from the last few days, where the web has little visual evidence and search results are cluttered with lookalikes. The authors call this setting 'Image Generation in the Wild'.

To measure progress, they built an automated pipeline that converts recent headlines into prompts and keeps only those whose real search results are highly confusing. The result is Savanna, a benchmark of 500 prompts across 16 subcategories. On Savanna, even harnesses built on frontier models retrieve a correct reference for only about half of the prompts.

The paper proposes Locate-Corroborate-Isolate: first use text search to pin down the target's identity, then verify candidate images against multiple sources, and finally crop a clean reference for the generator. Training the Ranger-8B model on 15,000 such trajectories gave a search score of 67.2, beating frontier-model harnesses, and it produced the best generation quality with every generator tested. The authors release the pipeline, benchmark, training data, and model.