HiDream-ai/ReCo-Bench

ReCo-Bench是一个基于VLLM的视频编辑质量评估基准数据集，旨在全面有效地评估视频编辑质量。数据集包含480个视频-指令对，均匀分布在四个任务（对象添加、移除、替换和视频风格化）中。所有源视频均来自Pexels视频平台。数据集还详细描述了评估指标，包括编辑准确性、视频自然性和视频质量三个主要维度，每个维度下又细分为三个子维度。

ReCo-Bench is a VLLM-based evaluation benchmark designed to comprehensively and effectively assess video editing quality. The dataset contains 480 video-instruction pairs evenly distributed across four tasks (i.e., object add, remove, replace, and video stylization). All source videos are collected from the Pexels video platform. The dataset also details evaluation metrics, including three major dimensions: edit accuracy, video naturalness, and video quality, each further divided into three sub-dimensions.