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

Divide and Control: Generation of Multiple Component Comic Illustrations with Diffusion Models Based on Regression

  • Zixuan Wang,
  • Peng Du,
  • Zhenghui Xu,
  • Qihan Hu,
  • Hao Zeng,
  • Youbing Zhao,
  • Hao Xie,
  • Tongqing Ma,
  • Shengyou Lin

摘要

Diffusion-based text-to-image generation has achieved huge success in creative image generation and editing applications. However, when applied to comic illustrations, it still struggles to deliver predictable high-quality productions with multiple characters due to the interference of the text prompts. In this paper, we propose a practicable method to use ControlNet and stable diffusion to generate controllable outputs of multiple components. The method first generates images for individual components separately and then degenerates those images to a regressed form, such as line drawings or Canny edges. Those regressed forms of individual components are then merged and fed into ControlNet to generate the final image. Experiments show that this method is highly controllable and can produce high-quality comic illustrations with multiple components.