Scholarlink

<p>Generating fashion designs from sketches and textures can significantly enhance design efficiency. While existing image generation frameworks can produce desired images based on language or sketch prompts, generating fine-grained fashion designs, especially with intricate sketches and textures, remains challenging. This necessitates models that can comprehend features at various levels within sketches and texture images. In this work, we propose FashionBBDM, a controllable fashion design framework built upon the Brownian Bridge Diffusion Model. FashionBBDM generates fashion designs from either existing or synthesized textures without relying on textual prompts. Our model employs a pre-trained UNet as the backbone network and introduces a latent space sketch encoder to extract multi-scale features from conditional images. Unlike traditional diffusion models, our image generation process commences from conditional images rather than Gaussian noise. In each generation iteration, the encoder output is integrated with the backbone network features to predictively output the next step, culminating in high-quality fashion design images. We quantitatively and qualitatively evaluate FashionBBDM using clothing and shoe datasets, demonstrating its superior conditional generation performance compared to state-of-the-art methods. Ablation experiments and result analyses further underscore the effectiveness of our framework’s various components. Code and trained models are available on <a href="https://github.com/wzm206/FashionBBDM">https://github.com/wzm206/FashionBBDM</a>.</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

Controllable fashion rendering via brownian bridge diffusion model with latent sketch encoding

  • 期刊论文

  • 发表日期：

    2025年3月27日

查看全文
  • Zengmao Wang,
  • Jituo Li,
  • Wei Gao

摘要

Generating fashion designs from sketches and textures can significantly enhance design efficiency. While existing image generation frameworks can produce desired images based on language or sketch prompts, generating fine-grained fashion designs, especially with intricate sketches and textures, remains challenging. This necessitates models that can comprehend features at various levels within sketches and texture images. In this work, we propose FashionBBDM, a controllable fashion design framework built upon the Brownian Bridge Diffusion Model. FashionBBDM generates fashion designs from either existing or synthesized textures without relying on textual prompts. Our model employs a pre-trained UNet as the backbone network and introduces a latent space sketch encoder to extract multi-scale features from conditional images. Unlike traditional diffusion models, our image generation process commences from conditional images rather than Gaussian noise. In each generation iteration, the encoder output is integrated with the backbone network features to predictively output the next step, culminating in high-quality fashion design images. We quantitatively and qualitatively evaluate FashionBBDM using clothing and shoe datasets, demonstrating its superior conditional generation performance compared to state-of-the-art methods. Ablation experiments and result analyses further underscore the effectiveness of our framework’s various components. Code and trained models are available on https://github.com/wzm206/FashionBBDM.