Scholarlink

<p>Dunhuang murals suffer from fragile pigments and complex iconography due to aging and historical interventions, making culturally faithful digital restoration challenging. Existing methods rely on static emotion–style mappings, offer limited interactivity, and overlook Dunhuang-specific symbolic semantics. This study proposes an emotion-driven multimodal generative framework for mural synthesis and restoration. Spatial motif priors are extracted using 2D-FFT and CNNs to capture periodic ornament structures. Text and speech inputs are fused into a continuous emotion embedding that regulates interpretable style parameters. An asynchronous latent-space partitioning strategy enables localized editing while preserving structural coherence, and a Dunhuang Symbol Knowledge Graph constrains iconographic semantics. Experiments on an expert-annotated dataset demonstrate improved emotion–style alignment, cultural-symbol fidelity, and motif consistency. Results are evaluated using CLIP similarity, CSA, PCI, and FID. Seed-level verification records support recomputation of reported metrics, offering a traceable, culturally grounded approach for digital heritage generation.</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

Emotion-driven multimodal generation of Dunhuang murals with spatial motif priors for interactive digital heritage

  • 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
  • Mengmeng Yuan,
  • Hainan Zhou,
  • Siyu Liu

摘要

Dunhuang murals suffer from fragile pigments and complex iconography due to aging and historical interventions, making culturally faithful digital restoration challenging. Existing methods rely on static emotion–style mappings, offer limited interactivity, and overlook Dunhuang-specific symbolic semantics. This study proposes an emotion-driven multimodal generative framework for mural synthesis and restoration. Spatial motif priors are extracted using 2D-FFT and CNNs to capture periodic ornament structures. Text and speech inputs are fused into a continuous emotion embedding that regulates interpretable style parameters. An asynchronous latent-space partitioning strategy enables localized editing while preserving structural coherence, and a Dunhuang Symbol Knowledge Graph constrains iconographic semantics. Experiments on an expert-annotated dataset demonstrate improved emotion–style alignment, cultural-symbol fidelity, and motif consistency. Results are evaluated using CLIP similarity, CSA, PCI, and FID. Seed-level verification records support recomputation of reported metrics, offering a traceable, culturally grounded approach for digital heritage generation.