MJ-BENCH-VIDEO
收藏资源简介:
MJ-BENCH-VIDEO是一个大规模的视频偏好数据集,由北卡罗来纳大学教堂山分校等机构创建。该数据集包含5个评价方面:对齐度、安全性、细腻度、连贯性与一致性、偏见与公平性,共计28个细粒度评价标准,旨在为视频生成模型的评价提供全面的基准。数据集通过三种策略收集了视频对及其对应的提示,经过筛选和注释后,形成了包含5421个数据条目的高质量数据集。该数据集适用于评估视频生成模型在多个方面的表现,推动更先进视频奖励模型的发展。
MJ-BENCH-VIDEO is a large-scale video preference dataset developed by institutions including the University of North Carolina at Chapel Hill. It encompasses five core evaluation dimensions: alignment, safety, detailedness, coherence and consistency, as well as bias and fairness, comprising a total of 28 fine-grained evaluation criteria, and is designed to provide a comprehensive benchmark for the evaluation of video generation models. The dataset collects video pairs and their corresponding prompts via three strategies, and after rigorous filtering and annotation, a high-quality dataset containing 5421 data entries is established. This dataset can be utilized to assess the performance of video generation models across multiple dimensions, thereby advancing the development of more advanced video reward models.




