Agentic Image Restoration Links Closed-Loop Controllers with Protocol-Aware Evaluation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Image restoration is shifting from fixed inference pipelines towards systems that diagnose degradation, choose tools, assess intermediate results, and revise their actions. However, the term “agentic” currently spans architectures with different control mechanisms and evaluation assumptions. We review representative studies published from 2018 to 2026 and offer a controller-centred synthesis of agentic image restoration (AIR). Our perspective separates four functional components, perception, decision, action, and reflection or memory, from five controller families: reinforcement learning, prompt conditioning, multimodal large language model tool use, memory-augmented planning, and multi-agent coordination. Evidence from photography, autonomous driving, medical reconstruction, remote sensing, and microscopy shows that controller comparisons are often confounded by incompatible degradations, tool pools, feedback signals, and budgets. We therefore pair the taxonomy with protocol-aware reporting priorities for output quality, downstream utility, execution cost, failure recovery, and provenance. This framework clarifies the boundary between conditioned restoration and autonomous control, identifies unresolved evaluation gaps, and provides a practical map for developing reliable and auditable AIR systems.