Frequency Domain Disentanglement for Arbitrary Neural Style Transfer

Arbitrary neural style transfer has been a popular research topic due to its rich application scenarios. Effective disentanglement of content and is the critical factor for synthesizing an image with arbitrary style. The existing methods focus on disentangling feature representations in spatial domain where components are innately entangled difficult be disentangled clearly. Therefore, these always suffer from low-quality results because sub-optimal disentanglement. To address such challenge, this paper proposes frequency mixer (FreMixer) module that disentangles re-entangles spectrum domain. Since have different frequency-domain characteristics (frequency bands patterns), FreMixer could well disentangle two components. Based module, we design novel Frequency Domain Disentanglement (FDD) framework transfer. Qualitative quantitative experiments verify proposed method can render better stylized compared state-of-the-art methods.