Scholarlink

<p>Custom text-to-face generation aims to generate images based on user input prompt and reference image that both match the textual description and remain identity consistent. Currently, some methods optimize both scene and character generation through multitask learning, or focus only on facial features, thus neglecting the importance of separating and accurately modeling character description (e.g., “young woman wearing sunglasses”) and scenarioized description (e.g., “sunbathing at the beach”) in prompt, resulting in the model is not accurate enough in scene restoration and character performance. To address these issues, we propose CustomFace, a novel generation framework that ensures identity consistency while realizing scene diversity. Firstly, fine-grained identity extractor is proposed to focus on the extraction of multi-scale identity features. Secondly, the instance-level cross-attention learning strategy refines the scene and character regions for instance-level modeling, effectively solving the scene–character decoupling problem. Extensive experimental results demonstrate that our method outperforms the state of the art in identity accuracy and scene 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

Instance-level cross-attention learning for fine-grained customizable face generation

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Xueping Wang,
  • Yixuan Gao,
  • Hang Li,
  • Feihu Yan,
  • Guangzhe Zhao

摘要

Custom text-to-face generation aims to generate images based on user input prompt and reference image that both match the textual description and remain identity consistent. Currently, some methods optimize both scene and character generation through multitask learning, or focus only on facial features, thus neglecting the importance of separating and accurately modeling character description (e.g., “young woman wearing sunglasses”) and scenarioized description (e.g., “sunbathing at the beach”) in prompt, resulting in the model is not accurate enough in scene restoration and character performance. To address these issues, we propose CustomFace, a novel generation framework that ensures identity consistency while realizing scene diversity. Firstly, fine-grained identity extractor is proposed to focus on the extraction of multi-scale identity features. Secondly, the instance-level cross-attention learning strategy refines the scene and character regions for instance-level modeling, effectively solving the scene–character decoupling problem. Extensive experimental results demonstrate that our method outperforms the state of the art in identity accuracy and scene restoration.