A Comprehensive Survey on Multi-Agent Systems for Audio-Visual Generation and Understanding

作者：Wentao Lei, Yan Rong, Jinting Wang, Tianxin Xie, Guanjie Huang, Li Liu · 年份：2025 · DOI：10.36227/techrxiv.176583352.26143196/v1 · 研究领域：Music Technology and Sound Studies、Multimodal Machine Learning Applications、Generative Adversarial Networks and Image Synthesis

The rapid evolution of artificial intelligence for multimedia has progressed from single-modality tasks to complex audiovisual generation and understanding, exposing fundamental limitations in end-to-end paradigms. These single-model approaches, often operating as black boxes, struggle with consistency, controllability, and robustness when tasked with sophisticated creative projects or complex analytical tasks. To address these challenges, Multi-Agent Systems (MAS) are emerging as a transformative paradigm, decomposing complex problems into collaborative systems of specialized agents that mirror human creative workflow. This survey provides the first comprehensive review of MAS for both audiovisual generation and understanding. We begin by detailing the foundations of MAS, including the core mechanisms of agent action, communication, and coordination. Building on this, we establish distinct taxonomies for both generation and understanding, organizing existing systems by dominant modality while examining key applications from creative industries to complex perception tasks. Furthermore, we offer a holistic evaluation ecosystem, summarizing key datasets and categorizing evaluation metrics from low-level fidelity to high-level agent reasoning. Finally, we identify critical challenges and chart future research directions toward unified architectures, world model integration, and enhanced human-AI collaboration. This work aims to provide a comprehensive view for researchers in thi...