LVD-2M
收藏资源简介:
LVD-2M是由香港大学和字节跳动合作创建的长视频数据集,旨在促进长时间视频生成模型的研究。该数据集包含200万条长时间视频,每条视频时长超过10秒,且无场景切换,具有丰富的动态内容和时间密集的描述。数据集的创建过程包括从多个开源数据集中筛选高质量的长视频,并使用分层视频描述方法生成时间密集的描述。LVD-2M的应用领域主要集中在AI辅助的电影制作等需要长时间视频生成的场景,旨在解决现有数据集在长时间视频生成方面的不足。
LVD-2M is a long-duration video dataset jointly created by The University of Hong Kong and ByteDance, aiming to advance research on long-form video generation models. This dataset contains 2 million long videos, each with a duration exceeding 10 seconds, no scene cuts, rich dynamic content, and temporally dense descriptions. The dataset construction process involves screening high-quality long videos from multiple open-source datasets, and generating temporally dense descriptions using a hierarchical video description method. The primary application scenarios of LVD-2M focus on AI-assisted filmmaking and other scenarios requiring long-duration video generation, aiming to address the shortcomings of existing datasets in long-form video generation.

- 1LVD-2M: A Long-take Video Dataset with Temporally Dense Captions香港大学 · 2024年



