Scholarlink

<p>Diffusion models have shown strong potential for image restoration, but their use in historical textile pattern restoration remains limited. Using a Tang-dynasty kesi hunting-pattern fragment unearthed in Xinjiang as a case study, this paper proposes a human–AI restoration method based on the Weavega Diffusion Model (WDM). The method combines a heritage-oriented pattern dataset, a damage-aware workflow, conditional WDM generation, retrieval-augmented generation (RAG) for severely damaged regions, protocol-based expert screening, and optional VAE refinement. Experimental results show that, under the same sampling budget, the RAG + WDM candidate pool showed improved expert-evaluation results compared with the WDM-only pool in the present case-based comparison, with the mean weighted score rising from 3.76 to 4.06, the median from 3.50 to 4.25, and the proportion of candidates scoring 4.25 or above from 22.2% to 55.6%. The study offers a practical pathway for human–AI collaboration in historical textile pattern restoration.</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

Weavega: a deep learning-based “human + AI” Method for the Restoration of Tang Dynasty Hunting-pattern Kesi Silk Fragments

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月29日

查看全文
  • Huan Wang,
  • Dong Han,
  • DengXu Cai,
  • HaoWen Xu,
  • KaiBing Zhang

摘要

Diffusion models have shown strong potential for image restoration, but their use in historical textile pattern restoration remains limited. Using a Tang-dynasty kesi hunting-pattern fragment unearthed in Xinjiang as a case study, this paper proposes a human–AI restoration method based on the Weavega Diffusion Model (WDM). The method combines a heritage-oriented pattern dataset, a damage-aware workflow, conditional WDM generation, retrieval-augmented generation (RAG) for severely damaged regions, protocol-based expert screening, and optional VAE refinement. Experimental results show that, under the same sampling budget, the RAG + WDM candidate pool showed improved expert-evaluation results compared with the WDM-only pool in the present case-based comparison, with the mean weighted score rising from 3.76 to 4.06, the median from 3.50 to 4.25, and the proportion of candidates scoring 4.25 or above from 22.2% to 55.6%. The study offers a practical pathway for human–AI collaboration in historical textile pattern restoration.