Skip to content

Dataset · Soul-AILab

330k talking head clips in 15+ languages

Soul-AILab/VividHead

782 hours of head-cropped speech videos: 330k clips, 15+ languages, varying ages.

What's inside

A large talking head dataset: 330k clips and 782 hours of video cropped to the speaker's head. The material covers over 15 languages and a wide age range of people. Collected for the image-to-video task of real-time continuous streaming generation of a speaking person.

What you can do with it

  • Training image-to-video models for portrait animation from speech
  • Training talking-head systems with streaming generation
  • Evaluating multilingual talking head synthesis models
Tasks
Image → video
Size
100K–1M rows
Downloads this month
1.9K
Likes
69

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("Soul-AILab/VividHead")