A Unified Arbitrary Style Transfer Framework via Adaptive Contrastive Learning

This work presents Unified Contrastive Arbitrary Style Transfer (UCAST), a novel style representation learning and transfer framework, that can fit in most existing arbitrary image models, such as CNN-based, ViT-based, flow-based methods. As the key component tasks, suitable is essential to achieve satisfactory results. Existing approaches based on deep neural networks typically use second-order statistics generate output. However, these hand-crafted features computed from single cannot leverage information sufficiently, which leads artifacts local distortions inconsistency. To address issues, we learn directly large number of images contrastive by considering relationships between specific styles holistic distribution. Specifically, present an adaptive scheme for introducing input-dependent temperature. Our framework consists three components: parallel transfer, domain enhancement module effective distribution, generative network transfer. Qualitative quantitative evaluations show results our approach are superior those obtained via state-of-the-art The code available at https://github.com/zyxElsa/CAST_pytorch.