软件导刊2026，Vol.25Issue(8)：164-172,9.DOI:10.11907/rjdk.251343
一种多模态跨领域的风格迁移方法
A Multi-Modal Cross-Domain Style Transfer Method
郑美君 1刘宏 1徐大宏 1李希1
作者信息
- 1. 湖南师范大学 信息科学与工程学院,湖南 长沙 410081
- 折叠
摘要
Abstract
Style transfer aims to apply artistic style to original content to generate innovative works.This paper proposes a multi-modal cross domain style transfer method based on the fusion of Vision Transformer(ViT)and Diffusion Transformer(DiT)to address the issues of insuffi-cient category attention and weak cross domain adaptability in multi-modal style transfer.Firstly,a unified framework is constructed by concat-enating ViT and DiT to enhance the collaborative processing capability of the model for text or image modal inputs.DiT's deep and label exten-sions are utilized to achieve cross modal feature alignment and solve the problem of modal bias in multi class style transfer.Secondly,improve the self attention mechanism of ViT by decoupling content and style features in the deep feature space,and achieve cross domain style adapta-tion through a parameterized fusion module.The experiment shows that the proposed method significantly improves the diversity and artistry of cross domain style transfer while preserving content semantics.Both quantitative indicators and qualitative comparisons are superior to existing methods,verifying its effectiveness in multi-modal cross domain scenarios.关键词
风格迁移/多模态/跨领域/视觉Transformer/扩散TransformerKey words
style transfer/multi-modal/cross-domain/Vision Transformer/Diffusion Transformer分类
信息技术与安全科学引用本文复制引用
郑美君,刘宏,徐大宏,李希..一种多模态跨领域的风格迁移方法[J].软件导刊,2026,25(8):164-172,9.
