Multi-Agent Image Restoration

TL;DR

MAIR tackles complex real-world image restoration by introducing a degradation-prior that categorizes degradations into scene, imaging, and compression, and reversing them in the inverse order with a three-stage restoration. It adopts a two-level multi-agent system, with a scheduler for planning and multiple experts for degradation-specific tool applications, guided by a tool registry and LLM-based reasoning. Empirical results on synthetic and real-world datasets show competitive reconstruction quality and improved efficiency over prior agentic methods, along with strong extensibility and instruction-following capabilities. The work advances practical IR by combining principled degradation ordering, modular tools, and training-free planning to reduce search complexity and computational cost.

Abstract

Image restoration (IR) is challenging due to the complexity of real-world degradations. While many specialized and all-in-one IR models have been developed, they fail to effectively handle complex, mixed degradations. Recent agentic methods RestoreAgent and AgenticIR leverage intelligent, autonomous workflows to alleviate this issue, yet they suffer from suboptimal results and inefficiency due to their resource-intensive finetunings, and ineffective searches and tool execution trials for satisfactory outputs. In this paper, we propose MAIR, a novel Multi-Agent approach for complex IR problems. We introduce a real-world degradation prior, categorizing degradations into three types: (1) scene, (2) imaging, and (3) compression, which are observed to occur sequentially in real world, and reverse them in the opposite order. Built upon this three-stage restoration framework, MAIR emulates a team of collaborative human specialists, including a "scheduler" for overall planning and multiple "experts" dedicated to specific degradations. This design minimizes search space and trial efforts, improving image quality while reducing inference costs. In addition, a registry mechanism is introduced to enable easy integration of new tools. Experiments on both synthetic and real-world datasets show that proposed MAIR achieves competitive performance and improved efficiency over the previous agentic IR system. Code and models will be made available.

Figures (7)

• Figure 1: Comparison between our proposed MAIR and typical agentic IR approaches.(a) The state-of-the-art method AgenticIR agenticir employs a single agent for perception, planning, restoration, etc., suffering from resource-intensive searches and trials for degradation removal. (b) We decompose the complex IR problem into manageable sub-tasks and address them using multiple collaborative agents, under our proposed three-stage restoration framework, achieving improved performance, efficiency, and flexibility.
• Figure 2: Illustration of the inference workflow of MAIR.(Left) Given an LQ image and a user instruction, a "scheduler" agent first obtains the coarse perception results of degradation types using DepictQA. It then inputs the experience, perception results, and user instruction into GPT-4o to formulate an overall restoration plan, following our three-stage framework. (Right) A group of "expert" agents sequentially removes degradations and outputs the reconstructed result, adhering to scheduler's plan. Each expert specializes in a single degradation and uses GPT-4o to intelligently select and apply a list of candidate tools to current image, effectively removing degradation based on the image, instruction, DepictQA's fine-grained perception results of degradation levels, tool registry information, and reflection.
• Figure 3: Qualitative comparison on three images from real-world paired (top), unpaired (middle) datasets, and Group A agenticir (bottom).
• Figure 4: Ablation study of three-stage framework on two images from real-world paired (left) and LHP-/Real-Rain-Val (right) datasets.
• Figure 5: Ablation study of two-level multi-agent system design regarding the selection of appropriate tools for effective noise removal on an image from Group B (left), and experience-guided formulation of restoration plans on a synthesized LQ image from MiO100 (right).