COTA: Color-guided Optimal Transport Attention for Arbitrary Style Transfer

: Arbitrary style transfer seeks to preserve the semantic structure of a content image while transferring the visual appearance of an arbitrary reference style image. Existing attention-based methods usually establish content-style correspondences through pairwise feature similarity, which can become unreliable when the two images differ substantially in semantics, texture, or color composition. We therefore propose COTA, a Color-guided Optimal Transport Attention framework that complements scaled dot-product attention with a structured color-correspondence prior. A shared ColorMaskNet performs soft color decomposition in the Lab color space and extracts representative color prototypes and component proportions from both images. A learnable color metric and Sinkhorn optimal transport then estimate a globally coupled soft transport plan subject to the color distributions of the two images. In contrast to hard assignment or independently normalized color matching, this formulation supports flexible one-to-many and many-to-many correspondences while coordinating color allocation globally. The transport plan is projected into layer-wise spatial priors and integrated with feature affinity in a multi-scale gated-attention framework. Experiments show that COTA provides competitive content preservation and style rendering, with coherent structures, stable local stylization, and consistent regional color allocation.