Real-world text-to-image requests are often underspecified. Qwen-Image-Agent treats the prompt as partial context and builds the full generation context before generating: planning identifies what is missing, and separate modules gather it via reasoning, web search, memory of past turns, and feedback. It is training-free and works with existing image generators.
The system is evaluated with a new benchmark, IA-Bench, covering Plan, Reason, Search, and Memory across 17 real-world tasks, 730 test instances, and 1801 checklist items. It outperforms strong agentic baselines on IA-Bench, Mindbench, and WISE-Verified. The interesting point is that the bottleneck in practical image generation is not rendering quality but gathering the right context.