In order to solve the problem of local artifacts and local feature loss in the output of unsupervised image style conversion model, an image style conversion model based on channel grouping attention mechanism was proposed. In the generator part of the model, a channel grouping attention residual block was used to enhance the extraction of image features and the utilization of effective features. In the discriminator part of the model, a dual discriminator structure was adopted, the added local discriminator was used to enhance the identification of the generated image details, and the multi-resolution global discriminator was used to enhance the content rationality and structural coherence of the generated image. The experimental results show that the unsupervised model not only has smaller volume, but also can obtain higher NIMA aesthetic score and LPIPS diversity score than other methods such as BicycleGAN and MUNIT. The model also performs well in the graphic design migration application task of packaging products.