Supplementary Material of ET-NET: Error Transition Network for Arbitrary Style Transfer

We assemble the non-local block into the bottleneck layer of the proposed framework to capture the long-range dependencies between pixels and make the processed error features more compatible to the current stylized results. The architecture of the revised non-local block is shown in Figure 1. Different from the way used in [8], where all the inputs come from the same image, we feed the top feature of stylized image f in and the full error feature ∆E 4 into the block. Then we measure the similarities between the error feature of one pixel and the features for stylized result at other locations to determine what error information to be transited from one pixel to another. In this way, we find that it is effective to capture long-range dependency with the advantages in respecting texture consistency.