Figure 2.

FaceMotionPreserve has a generator module that takes an original face and source identity as input, and outputs a de-identified face. During training, FaceMotionPreserve employs a discriminator module to improve image quality and preserve high semantic attributes, and a landmark extractor module to enhance facial dynamics preservation. Different losses are involved: identity loss ($Los{s}_{\mathit{ID}}$) enforces identity replacement, image loss ($Los{s}_{\mathit{img}}$) and discriminator’s GAN loss ($Los{s}_{\mathit{GAN}}$ and $Los{s}_{\mathit{GP}}$) and feature loss ($Los{s}_{\mathit{fea}}$) improve image reconstruction, and landmark loss ($Los{s}_{\mathit{lmk}}$) enhances facial movement preservation.