Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces EvoIR-Agent, a self-evolving image restoration agentic system designed to overcome the limitations of existing MLLM-driven approaches in degradation coupling scenarios. It addresses the trade-off between training-based (efficient but inflexible) and training-free (flexible but inefficient) methods by formulating explicit experience components and organizing them into a hierarchical pool. This pool is then updated through a self-evolving mechanism using accumulated records, leading to improved performance and efficiency in image restoration tasks.
Achieve Pareto-optimal image restoration by letting an agent learn from its own experience, striking a balance between performance and efficiency that surpasses existing methods.
Multimodal Large Language Model (MLLM)-driven image restoration agent demonstrates effectiveness in degradation coupling scenarios by flexibly selecting tools and determining removal orders. However, their zero-shot planning often fails without experience, necessitating severe trial-and-error overhead to achieve satisfactory outcomes. Currently, two paradigms are employed to address this issue, yet a dilemma persists: Training-based methods embed intrinsic experience into parameters, achieving high inference efficiency but lacking compatibility with new tools or degradation. In contrast, training-free methods utilize explicit experience storage for compatibility but still incur trial-and-error overhead due to naive experience. To resolve the dilemma, we propose EvoIR-Agent, which first systematically formulates the experience components of a training-free image restoration agent. Subsequently, a hierarchical experience pool is constructed, which enables coarse-to-fine guidance for diverse tools and removal orders. Furthermore, a self-evolving mechanism is introduced to update the pool from scratch using accumulated records, thereby greatly improving performance and efficiency. Extensive experiments reveal that EvoIR-Agent achieves a significant lead in the full reference metrics and yields a remarkable Pareto-optimal balance between performance and efficiency compared to the state-of-the-art methods.