- [2026/10] 📄 Paper is now available on arXiv.
- [Upcoming] 🔓 Code, Savanna benchmark, Ranger-15K dataset, and Ranger-8B weights are undergoing internal open-source review and will be released here upon approval.
Agentic image generation equips image generators with web search agents to handle entities beyond their training data. Yet real-world requests often target entities that appeared only days ago—a setting we call Image Generation in the Wild, which poses three core challenges:
- Temporal Novelty: Newly introduced entities lack stable visual records on the web.
- Semantic Crowding: Search results are dominated by look-alike predecessors, competitors, or related entities.
- Spatial Entanglement: Target entities rarely appear in isolation, often mixed with other objects or cluttered backgrounds that mislead downstream generators.
To continually construct time-sensitive, high-confusion prompts without relying on static entity lists, we build a four-stage automated pipeline from live news:
- World Snapshot Collection: Collect recent news and supplementary reports, restricting all subsequent searches to ≤ 3 days after publication to simulate temporal novelty.
- Entity Extraction: Extract target entities either directly from text (text-oriented) or from secondary objects visible only in retrieved photos (image-oriented, introducing spatial entanglement).
- Confusion Filtering & Verification: Retain only queries with retrieval confusion κ(u) ≥ 0.5 (semantic crowding) and cross-validate positive images to guarantee a single unambiguous ground truth.
- Prompt Assembly: Crop a clean ground-truth reference and compose a natural generation prompt paired with evaluation rubrics.
Built with this pipeline from July 2026 news and verified by three independent annotators, Savanna comprises 500 prompts across 16 subcategories in 5 domains (Entertainment, Sports, Fashion, Products, Culture). Even the strongest frontier-model agent harness reaches only 51.5 in search accuracy, and end-to-end search-enabled generators peak at 34.9 in generation score.
Built on Qwen3-VL-8B-Instruct, Ranger-8B follows the Locate–Corroborate–Isolate paradigm to deliver a grounded prompt and clean reference images to any downstream generator:
- Locate: Pin down the target's exact identity via text search and webpage reading before querying images.
- Corroborate: Cross-verify candidate images against gathered textual facts, source contexts, and each other.
- Isolate: Crop the verified target to remove distracting objects and backgrounds.
Ranger-8B is trained on Ranger-15K (15,119 verified trajectories synthesized from Jan–Jun 2026 news, strictly prior to Savanna's evaluation window).
Search score measures reference retrieval accuracy; generation score evaluates target identity and prompt fidelity across two downstream generators.
| Agent | Backbone | Search | Gen. (Qwen-Image-Edit-2511) | Gen. (Nano Banana Pro) |
|---|---|---|---|---|
| Mind-Brush | GPT-5.1 | 35.8 | 29.5 | 42.2 |
| GenEvolve | Qwen3-VL-8B | 37.5 | 34.2 | 42.7 |
| Gen-Searcher | Qwen3-VL-8B | 43.5 | 37.9 | 47.3 |
| GenClaw | Claude-Opus-5 | 49.8 | 36.5 | 46.4 |
| SCOPE | GPT-6-Astra | 51.5 | 37.3 | 47.6 |
| Ranger-8B (Ours) | Qwen3-VL-8B | 67.2 (+15.7) | 46.8 (+8.9) | 58.9 (+11.3) |
Using Qwen-Image-Edit-2511 as the downstream generator, Ranger-8B transfers zero-shot to established benchmarks and outperforms prior agentic methods:
| Agent | Backbone | Mind-Bench (Overall) | KnowGen (K-Score) |
|---|---|---|---|
| Mind-Brush | GPT-5.1 | 15.8 | 16.74 |
| GenEvolve | Qwen3-VL-8B | 18.4 | 22.91 |
| Gen-Searcher | Qwen3-VL-8B | 19.8 | 26.35 |
| Ranger-8B (Ours) | Qwen3-VL-8B | 28.2 (+8.4) | 27.75 (+1.40) |
Uncropped outputs generated by Nano Banana Pro conditioned on prompts and references retrieved by Ranger-8B across all five Savanna domains:
@article{li2026agentic,
title={Agentic Image Generation in the Wild},
author={Li, Yifan and Qin, Bowen and Zhou, Kun and Zhao, Haochen and Liu, Zikang and Li, Zheng and Hao, Minsheng and Cheng, Yuhao and Zou, Haosheng and Chen, Jeff and Zhao, Wayne Xin},
journal={arXiv preprint arXiv:2610.23269},
year={2026}
}