Towards mitigating uncann(eye)ness in face swaps via gaze-centric loss terms

Ethan Wilson, Frederick Shic, Sophie Jörg, Eakta Jain

Abstract: Highlights•Face swaps are generally perceived as more uncanny than real faces.•Pretrained gaze estimation models can be used to design targeted gaze loss terms.•Targeted loss equations decrease gaze angle error of generated face swaps.•Our proposed loss term makes the eyes less prominent in human deepfake detection.