Step-Video-TI2V-Eval
收藏资源简介:
Step-Video-TI2V-Eval是一个针对文本驱动的图像到视频生成任务的新基准数据集,由StepFun创建。该数据集包含178个现实世界和120个动画风格的提示-图像对,旨在覆盖多样化的用户场景。数据集根据类别特定的属性进行精细分类,包括动态艺术元素、运动学元素等,以实现全面的表征。该数据集支持对生成的视频在指令遵循、主体和背景一致性以及物理定律遵守等方面的评估,为相关研究提供了基础。
Step-Video-TI2V-Eval is a novel benchmark dataset for text-driven image-to-video generation tasks, created by StepFun. This dataset comprises 178 real-world and 120 animation-style prompt-image pairs, intended to cover diverse user application scenarios. It is finely classified based on category-specific attributes including dynamic artistic elements, kinematic elements, and others, to enable comprehensive characterization of the generation task. This dataset supports the evaluation of generated videos across multiple dimensions such as instruction following, subject and background consistency, and compliance with physical laws, serving as a foundational resource for relevant research.




