cz-5f/LoVoRA

LoVoRA数据集是一个高质量基准数据集，专为文本引导、无需掩码的视频对象移除和添加而设计。该数据集基于高保真图像编辑对构建，并合成为连贯的视频序列，提供对齐的源-目标视频对、时空一致的时空掩码、详细的光流和遮挡图以及指令级编辑注释。数据集构建流程包括图像到视频转换、掩码生成、光流估计、掩码传播和高质量视频修复。

The LoVoRA Dataset is a high-quality benchmark designed for text-guided, mask-free video object removal and addition. Built from high-fidelity image editing pairs and synthesized into coherent video sequences, it provides aligned source–target video pairs, temporally consistent spatio-temporal masks, detailed optical flow and occlusion maps, and instruction-level editing annotations. The dataset is constructed through a robust pipeline including image-to-video translation, mask generation, optical flow estimation, mask propagation, and high-quality video inpainting.