OmniHuman can turn photos into realistic videos of people speaking, singing and moving naturally, based on 18,700 hours of human motion data.