Answers are grounded in this paper and cite their sources.
Image restoration, enhancement, editing, and controllable generation are increasingly performed by systems in which a controller diagnoses the image, decomposes the request, selects and configures tools, inspects intermediate results, and revises or stops. The word agent is applied to systems with very different degrees of runtime control, and the literature is spread across communities. This review defines agentic image processing by behavior: an identifiable controller uses observations of the input or of an intermediate image to choose later processing actions. We systematically search the literature and analyze representative systems in terms of task, controller, observation, action, feedback, memory, organization, training, and autonomy. We propose a four-axis taxonomy and an A0–A4 autonomy scale, describe a common perceive plan–act–verify–remember architecture, and compare restoration, retouching, editing, closed loop generation, workflow composition, and applications. Three quarters of these systems appeared in 2025 or 2026, driven by multimodal large language models that perceive images, plan, and call tools. Controllers are moving toward reinforcement learning, and nearly all systems rely on learned judges whose errors propagate through the loop. We introduce a six layer evaluation framework covering image quality, goal satisfaction, decisions, trajectories, efficiency, and human-centered reliability, a reporting checklist, design guidelines, and open problems.
No takes yet. Share an insight, caveat, or question.
Shangquan Sun (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: