Agentic Image Generation

Agentic image generation studies systems that plan and act over multiple steps instead of producing an image from a prompt in one pass. Common approaches use vision-language models as planners or critics, coordinate specialized agents, or call tools for web search, image editing, code-based layouts, and workflow routing; some are training-free, while others use supervised fine-tuning and reinforcement learning. Work targets challenges such as object counts, spatial relationships, factual details, personalization, and editable designs, with evaluations including GenEval, WISE, and task-specific benchmarks. A central tension is how to make iterative decisions and verification reliable without adding excessive latency, cost, or unwanted changes to parts of an image that are already correct.