From Task-Specific Models to Agentic Systems: Progress and Challenges of Agents in Image Restoration

Weixiong Sun, Xiang Yin, Chao Dong

Keywords: image restoration; low-level vision; agent; large language model;

TL;DR: This paper surveys the shift in image restoration from task-specific models to LLM-based agents, reviews their architectures and capabilities, and highlights key challenges for building more autonomous and reliable restoration systems.

Abstract:

As a cornerstone of low-level computer vision, image restoration encompasses tasks such as denoising, dehazing, and super-resolution, reconstructing high-fidelity visuals from multifaceted degradations. While traditional IR has transitioned from task-specialized architectures to unified general models, existing paradigms still grapple with fidelity and quality trade-offs in real-world restoration.
In recent years, agent technologies, particularly those driven by large language models, have brought a novel solution to IR, leveraging their strong capabilities in cross-modal understanding, general reasoning, and natural language interaction.
We present a systematic review of IR evolution, delineating its progression through three stages: task-specific models, general models, and agentic systems.
We provide a granular analysis of LLM-based cognitive architectures and their core technical underpinnings, further establishing a comprehensive intelligence-level standard for IR agent systems.
Finally, we scrutinize critical bottlenecks in efficiency, quality assessment, and ethical alignment, while envisioning future research frontiers such as autonomous evolution.
This survey serves as a theoretical and practical roadmap for the next generation of intelligent image restoration.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 6