The key challenge of eliminating artifacts in virtual reality （VR） multi-view images lies in their cross-scale complexity. The multi-scale generative adversarial network （GAN） adapts to the cross-scale characteristics of artifacts by introducing a multi-scale structure to hierarchically process the global structure and local details of images. To address this issue， a method for eliminating visual artifacts in VR multi-view images based on multi-scale GAN is proposed in this study. A multi-scale GAN consisting of a multi-scale generator and a multi-scale discriminator is designed. First， the DenseNet-121 network in the multi-scale generator is used to extract multi-scale features of VR multi-view images with artifacts. Then， the feature pyramid network （FPN） and convolutional block attention module （CBAM） are combined to realize the fusion and optimization of multi-scale features. After upsampling and residual connection processing， a basic artifact-removed image is generated. Furthermore， the multi-scale discriminator is adopted to distinguish the authenticity of the generated images from different scales， and the feedback information is sent to the generator for further correction of image artifacts. The experimental results show that the proposed method can effectively eliminate visual artifacts of different scales in VR multi-view images. When eliminating artifacts from multi-view images， its structural similarity， feature similarity and edge preservation factor are all no less than 0.96， and the perceptual similarity can be as low as 0.035 or below. This method can eliminate cross-scale artifacts simultaneously， take both global structure and local details into account， and thus effectively improve the overall quality of VR multi-view images.