Due to the obvious intra-class differences caused by camera perspective, many researchers begin to use GAN to expand data to maintain intra-class invariance. Nevertheless, the images generated by existing GAN have some defects such as blurred image and unreal background. To solve the above existing problems, a person re-identification algorithm that uses mixed coding of posture and appearance features is proposed in this paper. In the training phase, the image of the person is decomposed into posture features and appearance features, and the generated network can generate high-quality images by switching the appearance feature and posture feature and then mixing the features in the two images. The discriminant network feeds back the appearance characteristics of the generated image to the appearance coder of the generated network to achieve joint optimization, and uses multi-loss joint to further improve the quality of the generated image. In the testing phase, the network model is tested using the original data set. The rank-1/mAP on the Market-1501 and DukeMTMC-reID data sets can reach 93.4%/82.2% and 84.3%/70.5% respectively.