Human-Motion
收藏资源简介:
Human-Motion数据集是由华中科技大学和美团合作创建的,专门用于高保真人类视频生成任务。该数据集利用大型语言模型生成详细的动作描述,旨在解决现有数据集在身份保持和动作控制方面的不足。数据集的创建过程包括使用光学流估计器提取视频中的运动强度,并通过多层感知机生成运动嵌入。Human-Motion数据集主要应用于个性化视频生成领域,旨在提高视频生成中的身份一致性和动作控制的精细度。
The Human-Motion dataset was jointly developed by Huazhong University of Science and Technology and Meituan, and is specifically designed for high-fidelity human video generation tasks. It leverages large language models to generate detailed action descriptions, aiming to address the limitations of existing datasets in terms of identity preservation and motion control. The dataset creation pipeline includes extracting motion intensity from videos using optical flow estimators, and generating motion embeddings via multi-layer perceptrons (MLPs). The Human-Motion dataset is primarily applied in the field of personalized video generation, with the goal of improving identity consistency and the fine-grained precision of motion control in video generation.




