Scholarlink

<p>In recent years, diffusion models have shown significant potential in single image generation tasks. However, existing single image generation models based on diffusion models are not designed to train several images simultaneously and controllably generate a single image. To address this issue, we propose a method that learns a conditional diffusion model from multi-single images, which is named CDMM. As we illustrate, the conditional mechanism introduced by CDMM allows for joint training on multiple single images with just one model and enables controllable generation of single images. Moreover, through conditional interpolation, CDMM can achieve image morphing between training or generated images, thereby opening the door to multi-image manipulation tasks that original models cannot solve. Extensive experiments on a wide range of images and public datasets demonstrate that CDMM achieves a higher LPIPS than GAN-based methods and a lower SIFID than diffusion-based methods, indicating that CDMM can generate realistic and diverse images. Our code will be freely available for public use upon acceptance at GitHub. </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

Cdmm: learning a conditional diffusion model from multi-single images

  • 期刊论文

  • 发表日期：

    2026年2月28日

查看全文
  • Xianjie Zhang,
  • Min Li,
  • Yao Gou,
  • Yusen Zhang,
  • Yujie He

摘要

In recent years, diffusion models have shown significant potential in single image generation tasks. However, existing single image generation models based on diffusion models are not designed to train several images simultaneously and controllably generate a single image. To address this issue, we propose a method that learns a conditional diffusion model from multi-single images, which is named CDMM. As we illustrate, the conditional mechanism introduced by CDMM allows for joint training on multiple single images with just one model and enables controllable generation of single images. Moreover, through conditional interpolation, CDMM can achieve image morphing between training or generated images, thereby opening the door to multi-image manipulation tasks that original models cannot solve. Extensive experiments on a wide range of images and public datasets demonstrate that CDMM achieves a higher LPIPS than GAN-based methods and a lower SIFID than diffusion-based methods, indicating that CDMM can generate realistic and diverse images. Our code will be freely available for public use upon acceptance at GitHub.