Abstract: Currently, the realistic avatars converted from sketch avatars has the problems of insufficient lifelikeness and low face recognition rate. In response to this, a realistic avatar conversion method with improved CycleGAN is proposed. Firstly, a face feature extractor is added to the U-Net self-encoder. Secondly, the features extracted by the face feature extractor and the features in the U-Net decoder are fused using the channel connection method, and the fused features are further decoded. Finally, the base model CycleGAN is transformed into a supervised learning model so as to add image spatial loss and style loss to the converted avatar and the real avatar. The experimental results show that the improved model reduces the FID value by 27.31 and improves Rank-1 by 19 % on the CUHK test set, and reduces the FID value by 8.65 and improves Rank-1 by 4.1 % on the XM2VTS test set, compared with the results collected from the converted real avatar of the base model.