Animation Costume Style Migration Based on CycleGAN

Abstract

The traditional style migration of animation costumes is mainly performed between two paired animation costumes. However, the generalization ability is weak, and the migration effect is not good when the gap between the training and testing costumes is large. To address the above problems, this paper proposes a style migration method for animated costumes combining full convolutional network (FCN) and CycleGAN, which enables the instance style migration between animated costumes with specific targets. It is also verified that the training dataset is not the factor that causes the poor style migration of CycleGAN. The experiments demonstrate that the animation costume style migration method combining full convolutional network and CycleGAN increases the recognition ability and can achieve the local style migration of the animation costume while maintaining the integrity of the rest of the elements, and compared with CycleGAN, the method can effectively suppress the style migration in areas outside the target.