The existing unsupervised image translation methods have some problems in the process of translation, such as the excessive change of image’s content and the poor style transmission of the target image domain, which result in the poor quality and diversity of the generated image. In order to solve these problems, an Image-to-Image Translation method based on style disentangled and Adaptive Layer-Instance Normalization was proposed in this paper. Firstly, an auto-encoder based on disentangled representation was adapted as the generator of model, and the content encoder was shared to enhance the content consistency of image to image translation. Secondly, the Adaptive Layer-Instance Normalization was added to the decoder, to change the image style adaptively and improve the effect of style transmission. Thirdly, the random style code was integrated into the loss function, and the model was made to have the ability to generate random style images and enhance the diversity of generated images. Finally, this method was designed and implemented, then verified and analyzed it on multiple datasets. Compared with the existing unsupervised image translation methods, the experimental results show that the images generated by our method have higher quality and better diversity.