Scholarlink

<p>Diffusion models have advanced human animation generation, but existing methods suffer from insufficient facial micro-expression modeling and poor cross-frame identity consistency. To address these issues, we propose PIPC, a high-fidelity framework preserving identity and pose coherence. PIPC adopts a ReferenceNet-ArcFace-CLIP joint encoding system to extract texture, identity, and global semantic features from reference images. The Identity-Semantic Enhancer refines ArcFace identity features with CLIP semantics to mitigate identity drift, while the RefArcClip Fuser achieves effective local feature coupling via optimized attention and variance alignment. Additionally, the ID-Motion Joint Guidance Reconstructor integrates 3DMM to capture subtle expressions and resolve depth loss in traditional 2D keypoint-based methods. Experiments on TikTok and self-built FC-100 datasets show PIPC outperforms state-of-the-art methods in identity consistency and texture fidelity, validating its efficacy in high-quality identity-preserving human animation.</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

High-Fidelity Human Image Animation: Preserving Identity and Pose Consistency

  • 期刊论文

  • 发表日期：

    2026年2月19日

查看全文
  • Aizhe Wu,
  • Jing Guo,
  • Kai Wang,
  • Shuai Yang,
  • Yongzhen Ke

摘要

Diffusion models have advanced human animation generation, but existing methods suffer from insufficient facial micro-expression modeling and poor cross-frame identity consistency. To address these issues, we propose PIPC, a high-fidelity framework preserving identity and pose coherence. PIPC adopts a ReferenceNet-ArcFace-CLIP joint encoding system to extract texture, identity, and global semantic features from reference images. The Identity-Semantic Enhancer refines ArcFace identity features with CLIP semantics to mitigate identity drift, while the RefArcClip Fuser achieves effective local feature coupling via optimized attention and variance alignment. Additionally, the ID-Motion Joint Guidance Reconstructor integrates 3DMM to capture subtle expressions and resolve depth loss in traditional 2D keypoint-based methods. Experiments on TikTok and self-built FC-100 datasets show PIPC outperforms state-of-the-art methods in identity consistency and texture fidelity, validating its efficacy in high-quality identity-preserving human animation.