Scholarlink

<p>Generating videos under multimodal control conditions has been a significant challenge. In this paper, we propose X2Fashion, a novel approach for video generation guided by multimodal inputs. X2Fashion enables diverse and efficient motion video generation by leveraging pose sequences, source images, and motion text descriptions. This approach ensures that the character’s appearance remains consistent with the source image while adaptively adjusting its movements according to the semantics of the input text. To address challenges related to temporal continuity and the alignment of appearance with motion semantics, we propose a novel framework incorporating an appearance encoder, a pose encoder, and a temporal module. These components are designed to work synergistically, enabling the generation of temporally consistent videos that closely align with user intentions. Extensive experiments on the Fashion-text2video dataset demonstrate that X2Fashion outperforms existing methods, generating smooth and coherent fashion videos that accurately reflect the provided text semantics under the specified control conditions.</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

X2Fashion: temporally consistent fashion video generation guided by image, pose and text

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Yongjiang Xue,
  • Congwei Guo,
  • Aizhe Wu,
  • Yang Chen,
  • Yongzhen Ke,
  • Peirong Tang

摘要

Generating videos under multimodal control conditions has been a significant challenge. In this paper, we propose X2Fashion, a novel approach for video generation guided by multimodal inputs. X2Fashion enables diverse and efficient motion video generation by leveraging pose sequences, source images, and motion text descriptions. This approach ensures that the character’s appearance remains consistent with the source image while adaptively adjusting its movements according to the semantics of the input text. To address challenges related to temporal continuity and the alignment of appearance with motion semantics, we propose a novel framework incorporating an appearance encoder, a pose encoder, and a temporal module. These components are designed to work synergistically, enabling the generation of temporally consistent videos that closely align with user intentions. Extensive experiments on the Fashion-text2video dataset demonstrate that X2Fashion outperforms existing methods, generating smooth and coherent fashion videos that accurately reflect the provided text semantics under the specified control conditions.