Text prompts and videos generated using 5 popular Text-to-Video models plus quality metrics including user quality assessments

A collection of 201 prompts which are used to generate short-form videos using 5 popular text-to-video models namely Tune-a-Video, VideoFusion, Text-To-Vudeo Synthesis, Text2Video-Zero and Aphantasia. Each of the 1,005 generated videos is included along with automatically calculated quality metrics naturalness, text similarity between the original prompt and a generated text caption, and inception score, for each. Each video was rated by 24 different people and the data also includes the MOS scores for alignment between the generated videos and the original prompts, as well as for perception and overall quality of the video.Please cite this paper if using this dataset. GitHub URL for code for implementing video naturalness calculation is available at https://github.com/Chiviya01/Evaluating-Text-to-Video-Models

本数据集包含201条提示词（prompt），用于通过5款主流文本转视频模型生成短视频，这5款模型分别为Tune-a-Video、VideoFusion、Text-To-Vudeo Synthesis、Text2Video-Zero及Aphantasia。本次共生成1005条视频，每条视频均附带自动计算得到的三项质量指标：自然度、原始提示词与生成文本字幕间的文本相似度，以及起始评分（Inception Score）。所有视频均由24名受试者完成评分，数据集同时收录了生成视频与原始提示词的对齐度、视频感知质量及整体质量的平均意见得分（MOS）。使用本数据集时，请引用该论文。用于实现视频自然度计算的代码GitHub仓库地址为：https://github.com/Chiviya01/Evaluating-Text-to-Video-Models