MAGI Training Dataset
收藏arXiv2025-09-30 收录
下载链接：
https://github.com/SandAI-org/MAGI-1
下载链接
链接失效反馈官方服务：
资源简介：
该数据集是从数十PB原始视频和图片中精心挑选出的高质量视频片段，经过处理以确保视觉和运动质量，专门用于训练MAGI-1视频生成模型。该数据集包含了质量评估、降噪以及语义分布调整的筛选机制，确保其适用于自回归视频生成任务，规模达到了数十PB。
This dataset consists of high-quality video segments meticulously curated from tens of petabytes of raw video and image materials. It has been processed to guarantee optimal visual and motion quality, and is specifically designed for training the MAGI-1 video generation model. The dataset integrates screening mechanisms covering quality assessment, noise reduction, and semantic distribution adjustment, ensuring its applicability to autoregressive video generation tasks, with a total size of tens of petabytes.
提供机构：
MAGI-1 Research Team
