Microsoft Research Asia has released an AI model that can generate realistic, talking deepfake videos from a single still image and an audio track.
The model has been trained on
footage of approximately 6,000 talking faces
from the VoxCeleb2 dataset, and can animate
still images that lip-sync to a supplied
voice track, creating realistic vocal expressions
and natural head movements.
The technology, called
VASA-1, can reportedly generate synced videos at
512x512 pixels at 40 frames per
second without latency.
Photo credit: Microsoft Research Asia