You monitor the "Loss" value; as it drops, the face becomes clearer.

This is the video you want to put the face onto. High-bitrate, 1080p or 4K footage works best for tracking. 3. Extraction & Alignment The software must "find" the faces in every frame.

The industry standard for high-quality deepfakes. It requires thousands of images and hours (or days) of training a model.