Cross-modal style transfer guided by fine-grained text based on capsule network

This paper proposes a fine-grained text-guided cross-modal style transfer framework based on capsule networks, addressing the shortcomings of existing methods in multi-modal feature alignment, style decoupling, and fine-grained control. Firstly, a text-image collaborative encoder is constructed, using the dynamic routing capsule network to encode text attributes into a set of interpretable style-content capsules, achieving explicit separation of style and content. Secondly, a cross-modal attention style injection module is designed, using the routing coefficients between capsules to precisely map text style information to the image content representation, supporting fine-grained adjustment of local attributes such as color, texture, and brushstrokes. Finally, contrastive learning constraints are introduced to ensure consistency and authenticity of style transfer. Empirical results on public benchmarks demonstrate that the proposed approach achieves markedly better style controllability, content preservation, and visual fidelity than state-of-the-art competitors, opening a novel avenue for steerable cross-modal generation.