VideoReward
收藏资源简介:
VideoReward数据集是一个大规模的视频生成偏好数据集,由香港中文大学、清华大学、快手科技等机构联合创建。该数据集包含182,000条标注数据,涵盖了视觉质量(VQ)、运动质量(MQ)和文本对齐(TA)三个关键维度,旨在捕捉用户对生成视频的偏好。数据集通过12个先进的视频生成模型生成,并经过人工标注,标注过程包括对视频对的偏好选择。数据集的应用领域主要集中在视频生成模型的优化,旨在解决视频生成中的运动不连贯、与提示文本不对齐等问题。通过该数据集,研究人员可以训练多维度视频奖励模型,进一步提升视频生成的质量和用户满意度。
VideoReward Dataset is a large-scale video generation preference dataset jointly created by institutions including The Chinese University of Hong Kong, Tsinghua University, and Kuaishou Technology. It contains 182,000 annotated samples, covering three core dimensions: Visual Quality (VQ), Motion Quality (MQ), and Text Alignment (TA), aiming to capture user preferences for generated videos. The dataset is generated by 12 state-of-the-art video generation models and underwent human annotation, where the annotation process involves preference selection between video pairs. Its primary application focuses on optimizing video generation models, targeting prevalent issues in video generation such as incoherent motion and misalignment with prompt texts. With this dataset, researchers can train multi-dimensional video reward models to further improve the quality of video generation and user satisfaction.




