Multiangle feature fusion network for style transfer

Zhenshan Hu, Bin Ge, Chenxing Xia

Abstract: Highlights•We propose a new arbitrary style transfer network (MAFST), which can establish multiple angle associations between content features and style features, effectively solving the problem of repeated patterns in generated images.•We designed a Multi-Scale Style Capture module. It can capture shallow style features, and embed them into the generated image to optimize the stylization of the model.•We can achieve excellent video style transfer through fine-tuning and introducing patch-wise content temporal consistency loss. Our model can effectively reduce flickering and maintains content stability.