CanonSwap's "normative space" is its core innovative technology, which is essentially a methodological space for standardized processing of video frames. The core idea of the technology is to decouple the face information by deep learning algorithms intoprofilecap (a poem)Movement characteristicsTwo independent dimensions.
The specific realization process consists of three key steps:
- space transformation: Analyze the dynamic parameters such as head pose and expression of the original video frames using a motion extractor to uniformly map all faces to a standard forward pose
- feature separation: Facial identity information (e.g., shape of features, skin color, etc.) is stored completely separate from motion information (e.g., blinking, smiling, etc.) in canonical space
- Selective editing: Modify only the exterior feature vectors in canonical space, keeping the original motion parameters absolutely unchanged
This technology breaks the limitation that traditional face-swapping technology must synchronize the modification of appearance and motion, enabling the final generated video to simultaneously meet the dual standards of high fidelity of identity and zero dynamic distortion.
This answer comes from the articleCanonSwap: A tool for realizing high-fidelity face-swapping in videoThe
