global-optima-research/HDTF

HDTF数据集是一个为生成动态人头、视频配字和多模态虚拟人合成任务而专门整理和预处理的High-Definition Talking Face数据集版本。它包含了原始全长视频、81帧的短视频片段、使用OpenAI Whisper提取的音频嵌入、与视频片段对齐的棍状人姿态视频、多模态潜在张量（包括视频片段、姿态视频的潜在特征和字幕的文本嵌入）、每个片段的文本描述（字幕），以及训练、验证和测试视频片段的文件名列表。所有模态都通过一致的片段文件名对齐。

The HDTF Dataset is a curated and preprocessed version of High-Definition Talking Face, prepared for tasks such as talking-head generation, video captioning, and multimodal avatar synthesis. It includes original full-length videos, short video clips of 81 frames, audio embeddings extracted using OpenAI Whisper, stickman-style pose videos aligned with the clips, multimodal latent tensors (including latent features of clips, poses, and text embeddings from captions), textual descriptions (captions) for each clip, and lists of file names for training, validation, and test clips. All modalities are aligned via consistent clip filenames.