PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
March 21, 2026Journal of Image and Graphics0 citationsOpen Access

多模态引导裙装图像生成的结构化风格增强学习

MJMa JianiLLLi LiuFXFu Xiaodong

Key Points

  • To address redundancy and conflict in multimodal skirt image generation, a structured style enhancement learning method is proposed.
  • Input text descriptions to guide image generation.
  • Design dynamic attribute templates for seven key skirt attributes.
  • Implement a text-reversal semantic fusion mechanism.
  • Construct cross-domain image feature alignment with attention mechanisms.
  • Develop a dual-condition collaborative fusion framework for enhanced text and style representations.
  • Demonstrated improvements in Fréchet Inception Distance (FID) by 2.131 over competing methods.
  • Increased Learned Perceptual Image Patch Similarity (LPIPS) by 0.193 compared to other approaches.
  • Achieved a 17.57% improvement in Contrastive Language-Image Pre-training Score (CLIPScore).
  • Enhanced Texture Score (TS) by 8.29%, indicating superior image generation quality.

Abstract

目的针对多模态引导的裙装图像生成中存在的多角度文本注释信息冗余与冲突、跨区域风格传递能力有限以及语义与风格难以精细协同控制的问题,提出了一种结构化风格增强学习方法。方法以文本描述作为输入,针对裙装特点设计动态属性模板生成策略,智能提取并重构7类关键裙装属性,构建消除冗余与冲突的结构化文本提示;建立文本反转语义融合机制,将裙装图像特征经文本反转生成伪词嵌入,与结构化提示融合,形成语义丰富的文本表示;构建跨域图像特征对齐模块,引入跳跃交叉注意力,实现草图结构与风格图像的选择性融合并实现跨区域风格关联;建立双重条件协同融合框架,将增强的文本表示与跨域风格表示分层注入潜在扩散模型,精细控制语义与风格以生成裙装图像。结果实验在DressCode Multimodal数据集裙装子集上与目前较新的5种方法进行比较。结果表明,所提方法的弗雷歇起始距离(Fréchet inception distance,FID)和学习感知图像块相似度(learned perceptual image patch similarity,LPIPS)较对比方法提高2.131和0.193,对比语言图像预训练分数(contrastive language-image pre-training score,CLIPScore)和纹理分数(texture score,TS)分别提高17.57%和8.29%,说明本文方法具有更好的生成效果。结论本文提出的多模态引导裙装图像生成的结构化风格增强学习方法,能有效聚焦语义内容与风格结构间的深层关联,在确保多模态一致性的同时,实现高质量的裙装图像生成。

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Jiani et al. (2026) studied this question.

synapsesocial.com/papers/69be37dd6e48c4981c677da0https://doi.org/10.11834/jig.250338
Ask AI
Helpful
Bookmark
Share
View Full Paper