Hunan Embroidery Image Generation Method Based on Cross-Attention Conditional ControlNet
Bo Yang, Yukang Fu, Shunyao Sun, Ling Chen
Hunan Normal University
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Hunan embroidery image generation provides an efficient technical approach for the digital preservation and transmission of Intangible Cultural Heritage. However, existing artificial intelligence models suffer from insufficient structural constraint accuracy and poor stability when generating Hunan embroidery images with complex texture patterns. To address these limitations, this paper proposes a Hunan embroidery image generation model named Cross-Attention Conditional ControlNet(CACC).In CACC, the structural latent features produced by conditional ControlNet are redesigned from the conventional zero-convolution residual injection paradigm to a key–value injection mechanism at the cross-attention level. Specifically, structural edge feature maps are transformed into attention key–value pairs through a differentiable mapping function and are concatenated into the attention sequences of both double-stream and single-stream Transformer blocks. This design enables the control signals to participate directly in the attention weight computation stage, thereby influencing the attention distribution and achieving more expressive structure-guided generation. Experimental results demonstrate that, compared with the Flux-Kontext + ControlNet baseline, the proposed method improves the Peak Signal-to-Noise Ratio (PSNR) by 16.8%, reduces the Fréchet Inception Distance (FID) by 9.9%, and decreases the Learned Perceptual Image Patch Similarity (LPIPS) by 2.9%. A survey-based expert evaluation indicates that CACC demonstrates superior craftsmanship-related authenticity in stitch structure accuracy, material & luster fidelity, and regional style consistency. These improvements enhance structural preservation in high-frequency texture regions and improve style fidelity in generated Hunan embroidery images.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
计算机 / AIGenerative Adversarial Networks and Image Synthesis
Advanced Neural Network Applications · Advanced Image Processing Techniques