FuseFormer

What is FuseFormer? FuseFormer is a video inpainting model that uses a feedforward network to enhance subpatch level feature fusion. It is based on specialized Transformer-based technology with novel Soft Split and Soft Composition operations. These operations divide the feature map of a video into small patches and then stitch them back together. This enhances the video's overall quality by improving the fine-grained feature fusion of the video. How Does FuseFormer Work? FuseFormer works by