FDC-Swap: An efficient face swapping framework based on feature disentangling consistency

Face swapping images with high quality are critical for video production, privacy protection, and forgery detection, etc. In face swapping technology, one kernel element is the disentanglement of identity and attribute information. However, due to their deep entanglement in the latent space, effectively enhancing disentangling ability remains challenging. To tackle this challenge, this paper proposes the feature disentangling consistency: by executing the face swapping process (i.e., disentangling and recombining the identity and attribute information of two face images) in a pipeline twice, the generated twice-swapped images should be consistent with the original images. First, a new evaluation method (named FDC-Evaluation) is proposed to assess model performance according to disentangling consistency, which can be reflected in identity-attribute or image-level consistency. Meanwhile, to handle the limitation of missing ground truth, we introduce the FDC-Evaluation into original face swapping network, and propose a highly scalable and lightweight swapping framework (named FDC-Swap). Experiments demonstrate that the FDC-Evaluation overcomes traditional evaluation limitations, such as cognitive differences in identity-attribute features, identity assessment unfairness due to choices of identity encoders in training and evaluation, and operational complexity due to attribute feature diversity. Meanwhile, the FDC-Swap framework effectively enhances the performance of various existing face swapping networks. Code available at https://github.com/tzjoyzx/FDC_Swap.