SkyScript-100M
收藏资源简介:
SkyScript-100M是一个专注于短剧视频制作的大规模多模态数据集。该数据集汇集了来自互联网的6660部流行短剧,总计约80,000集,总时长超过2000小时,数据量高达10TB。通过关键帧提取与注释,SkyScript-100M提供了10亿对高质量的短剧本与拍摄脚本,为短剧视频生成领域提供了丰富的资源。数据集的构建过程包括多模态大语言模型的预注释、关键信息清洗、像素化处理以及后续的校准与优化,确保了数据的高质量与实用性。SkyScript-100M的应用领域广泛,旨在推动文本到视频的转换技术,解决短剧制作中的剧本优化问题,促进短剧视频生成领域的范式转变。
SkyScript-100M is a large-scale multimodal dataset dedicated to short-form video drama production. It compiles 6,660 popular short-form drama series sourced from the Internet, totaling approximately 80,000 episodes with an aggregate duration of over 2,000 hours and a total data volume of up to 10 TB. Through keyframe extraction and annotation, SkyScript-100M provides 1 billion high-quality pairs of short drama scripts and shooting scripts, serving as a rich resource for the field of short-form video drama generation. The dataset's construction workflow encompasses pre-annotation using multimodal large language models (LLMs), key information cleaning, pixelization processing, as well as subsequent calibration and optimization, which ensures the high quality and practical usability of the collected data. SkyScript-100M boasts a broad spectrum of application scenarios, with the goals of advancing text-to-video conversion technologies, resolving script optimization challenges in short-form drama production, and facilitating paradigm shifts in the field of short-form video generation.
SkyScript-100M: 1,000,000,000 Pairs of Scripts and Shooting Scripts for Short Drama
数据集概述
- 名称:SkyScript-100M
- 内容:包含1,000,000,000对短剧剧本和拍摄剧本的数据集
- 来源:从互联网收集的6,660个受欢迎的短剧剧集,每个剧集平均包含100个短剧集,总计约80,000个短剧集,总时长约2,000小时,总计10TB
- 处理:对每个剧集进行关键帧提取和标注,得到约10,000,000个拍摄剧本,并通过自研的大型短剧生成模型SkyReels进行100次剧本恢复
数据集用途
- 研究目的:基于SkyScript-100M,研究人员可以实现更深入和更远大的剧本优化目标,可能推动文本到视频领域的范式转变,并显著推进短剧视频生成领域的发展
相关资源
- 技术报告:Technical Report
- 生成模型:SkyReels
引用信息
bibtex @misc{tang2024skyscript100m, title={SkyScript-100M: 1,000,000,000 Pairs of Scripts and Shooting Scripts for Short Drama}, author={Jing Tang, Quanlu Jia, Yuqiang Xie, Zeyu Gong, Xiang Wen, Jiayi Zhang, Yalong Guo, Guibin Chen, Jiangping Yang}, year={2024}, eprint={2408.09333}, archivePrefix={arXiv}, primaryClass={cs.CL} }
联系信息
- 联系人:Jing Tang (唐晶)
- 邮箱:j_tang@hust.edu.cn




