Scholarlink

<p>Thangka artifact elements contain strict structures, intricate ornamentation, and rich cultural symbolism, posing challenges for controllable image generation. Existing text-to-image models often produce structural distortions, blurred details, and cultural semantic deviations. To address these issues, we propose a controllable diffusion-based method integrating cross-domain feature alignment and high-order interaction mechanisms. Built on a latent diffusion framework, the method introduces Frequency-Domain Cross-Normalization to align control and denoising features and Hierarchical High-Order Interaction Control to enhance long-range structural modeling and fine-grained ornamentation representation. We also construct TRA-15, a multimodal dataset of Thangka artifact elements. Experiments on TRA-15 demonstrate improved generation quality, structural accuracy, and ornamental detail preservation, while subjective evaluations indicate better controllability and visual-cultural consistency. Results on COCO further show moderate generalization ability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Controllable generation of Thangka artifact elements based on cross-domain feature alignment and high-order interaction mechanisms

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月8日

查看全文
  • Yanjiao Wei,
  • Tiejun Wang,
  • Lingmei Tao,
  • Bowen Liu

摘要

Thangka artifact elements contain strict structures, intricate ornamentation, and rich cultural symbolism, posing challenges for controllable image generation. Existing text-to-image models often produce structural distortions, blurred details, and cultural semantic deviations. To address these issues, we propose a controllable diffusion-based method integrating cross-domain feature alignment and high-order interaction mechanisms. Built on a latent diffusion framework, the method introduces Frequency-Domain Cross-Normalization to align control and denoising features and Hierarchical High-Order Interaction Control to enhance long-range structural modeling and fine-grained ornamentation representation. We also construct TRA-15, a multimodal dataset of Thangka artifact elements. Experiments on TRA-15 demonstrate improved generation quality, structural accuracy, and ornamental detail preservation, while subjective evaluations indicate better controllability and visual-cultural consistency. Results on COCO further show moderate generalization ability.