Abstract: Aimed at the low resolution and poor quality of face images generated by the DeepFakes face replacement method, a face replacement method based on self-attention generation confrontation network is proposed. The main body of the generation confrontation network adopted a U-shaped self-encoding symmetric structure to reduce the loss of feature information. We introduced the self-attention mechanism to better learn the texture characteristics of the image, improved the reconstruction quality of the generated image, and applied the Kalman filter to smooth the position of the bounding box on each frame thus reducing the face jitter. A comparative experiment was carried out on the FaceForensics++ dataset with the DeepFakes replacement method. The qualitative and quantitative experimental results prove that the method can better improve the quality of the generated image and reduce facial jitter.