OMG-Agent：解耦粗粒度到细粒度智能体工作流的鲁棒缺失模态生成

Multimodal learning frequently encounters missing modality challenges in real-world applications. Existing approaches, often relying on joint training or modality imputation, struggle to maintain robustness in complex scenarios. This paper introduces OMG-Agent, a novel framework for robust missing modality generation, characterized by its decoupled coarse-to-fine agentic workflows. OMG-Agent comprises two collaborative agents: a coarse-grained agent and a fine-grained agent. The coarse-grained agent initially leverages available modality information to generate a preliminary representation of the missing modality, capturing its primary features and global structure. Subsequently, the fine-grained agent iteratively refines this initial representation, incorporating multimodal information to progressively enhance the details of the missing modality, ensuring high generation quality and consistency with existing modalities. This decoupled design enables the model to effectively handle varying degrees of modality absence and significantly improves the robustness of the generated outputs. Experimental results demonstrate that OMG-Agent substantially outperforms state-of-the-art methods across multiple benchmark datasets, particularly excelling in addressing intricate missing patterns and boosting downstream task performance when utilizing the generated modalities. This work presents a promising new solution for the missing modality problem in multimodal learning.