Discussion on the Development of Face-swap Methods

Abstract

 Face-swap technology is a technology that can transfer source face identity information to the target face and maintain some other attributes of the target face, also known as deepfake. Among these methods, the first to appear were target-specific ones, which tend to have better performance but lower efficiency. As deep learning techniques continue to evolve, the pursuit of algorithmic generalization has begun, using pre-trained models to swap different faces. Although the generalized models are very efficient, the quality of certain face-swap images and videos generated is not satisfactory. Thus, people want to find an algorithm with both generality and performance. In the context of the current general pursuit of both generality and performance of deep forgery mehtods, this paper examines the development of deep forgery methods from the early days to the present and discusses the question of whether algorithm generality and performance can be achieved at the same time. After that, this paper proposes a deep forgery dataset, which contains the wild face videos we found from the Internet video community Bilibili, as well as the forged videos and images we generated using the four methods: Deepfakes, DeepFaceLab, FSGAN, and FaceDancer, and tests on this dataset the ability of the existing deep forgery methods for the generation capability of wild faces. Finally, we change the direction of identity vector usage for the effect of the target face on identity transfer during the generation process and propose a metric to measure the identity transfer error of the generated face caused by the target face.