370bfb31abd222b582245b977ea5f25a-AuthorFeedback.pdf

InFig.A,weshow4 the network architecture when only one example image is used (i.e.5 without the attention mechanism). This operation could be considered as a1x1 convolution with agroup size equal to the channel size. From this18 perspective, AdaIN is a constrained case of the proposed weight generation scheme, since our scheme can generate a19 convolutional kernel with group sizeequal to1andkernel sizelargerthan 1x1. Toour best knowledge, we are the first to show that weight generation can be used to solvethe adaptive36 video-to-video synthesis problem. Different videos arerandomly divided intotraining/test sets.