UniVA-Agent/UniVA-Bench

UniVA-Bench是一个用于评估智能视频系统的综合基准，包含四个类别的多步骤视频任务：理解长视频的美学和语义问题、基于不同输入生成视频、编辑长视频以保持叙事完整性和一致性、对具有多个实体和频繁遮挡的长视频进行分割。

UniVA-Bench is a comprehensive benchmark for evaluating intelligent video systems, consisting of four categories of multi-step video tasks: understanding aesthetics and semantics in long videos, generating videos based on various inputs, editing long videos to maintain narrative integrity and consistency, and segmenting long videos with multiple entities and frequent occlusions.