EVENT DOSSIER
On September 7, 2026, the arXiv cs.CV preprint platform published an academic paper titled AVENUE (Audio-Video Editing Understanding and Evaluation). This study proposes a unified framework aimed at handling audio and video editing tasks, understanding their semantic meanings, and conducting automated evaluations. AVENUE attempts to address the lack of unified standards and deep understanding in the field of audio and video generation and editing by integrating multimodal data with editing logic, providing a foundational reference for the standardized development of related technologies in the future.
研究人员推出 AVENUE（Audio-Video EditiNg Understanding and Evaluation），旨在解决音频 - 视频编辑中模型难以忠实保留非目标模态的问题。该框架包含两项核心贡献：一是涵盖 1,291 个源片段和 7,957 条编辑指令的基准数据集，覆盖音频、视频及音视频耦合三种类型；二是构建了一个样本特定且模态感知的评估体系，明确指定每个样本的预期修改与需保留内容。研究对联合、顺序及分离三种编辑范式下的代表性模型进行了系统分析，发现现有模型在编辑一种模态时，常会在另一种模态中引发非预期变化，无论其所属范式如何。AVENUE 数据集已公开于 Hugging Face。