Scholarlink

<p>Recent years have witnessed significant advances in deep learning for style transfer. Although diffusion models perform exceptionally well in image generation, they have limitations in modeling ink brushstrokes and capturing cultural meaning for Chinese ink painting. Current methods still face challenges including artifacts from inscriptions and seals, insufficient brushstroke details, unnatural ink diffusion, low-resolution output, and missing cultural elements. To solve these problems, we propose InkArtGAN—an end-to-end framework that improves detail handling, style consistency, image clarity and cultural features. Our method uses Hue-Saturation-Value (HSV) preprocessing to remove red seals and inscriptions, applies Visual Geometry Group 16-layer network (VGG-16)—based perceptual loss to refine brushstroke textures and ink diffusion, integrates Real-ESRGAN to improve image resolution, and adds customizable inscription and seal module. We compare visual quality, run quantitative tests, and have 10 art experts rate the images. Experimental results indicate that InkArtGAN outperforms baseline models (e.g., ChipGAN, CycleGAN) and Stable Diffusion in detail preservation, style consistency, clarity and cultural feature restoration. Ablation studies confirm each module’s contribution. Future work will improve ink fluidity through temporal modeling and expand multi-style training.</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

Inkartgan: a deep learning-based approach for digital generation of ink painting

  • 期刊论文

  • 发表日期：

    2025年8月20日

查看全文
  • XiaoNan Wang,
  • YiLe Du,
  • YiFei Pang

摘要

Recent years have witnessed significant advances in deep learning for style transfer. Although diffusion models perform exceptionally well in image generation, they have limitations in modeling ink brushstrokes and capturing cultural meaning for Chinese ink painting. Current methods still face challenges including artifacts from inscriptions and seals, insufficient brushstroke details, unnatural ink diffusion, low-resolution output, and missing cultural elements. To solve these problems, we propose InkArtGAN—an end-to-end framework that improves detail handling, style consistency, image clarity and cultural features. Our method uses Hue-Saturation-Value (HSV) preprocessing to remove red seals and inscriptions, applies Visual Geometry Group 16-layer network (VGG-16)—based perceptual loss to refine brushstroke textures and ink diffusion, integrates Real-ESRGAN to improve image resolution, and adds customizable inscription and seal module. We compare visual quality, run quantitative tests, and have 10 art experts rate the images. Experimental results indicate that InkArtGAN outperforms baseline models (e.g., ChipGAN, CycleGAN) and Stable Diffusion in detail preservation, style consistency, clarity and cultural feature restoration. Ablation studies confirm each module’s contribution. Future work will improve ink fluidity through temporal modeling and expand multi-style training.