Scholarlink

<p>Long-horizon alignment of musical structure with biomechanically valid motion remains elusive for 3D dance generation: existing diffusion models drift across phrases and produce foot-skate artifacts. We introduce PF–MambaDance, a conditional diffusion framework that factorises temporal and spatial modelling. A linear-time Mamba state-space backbone captures phrase-level dynamics, an anatomical graph-CNN enforces skeletal coupling, and multi-scale FiLM fuses beat/phrase cues. A Pose-Fusion prior (2D reprojection cycle with a frozen OpenPose-CNN-GCN) and lightweight physics regularisers jointly constrain bone lengths, joint limits and ground contacts, while a music<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\leftrightarrow \)</EquationSource> </InlineEquation>dance cycle head with lag-robust synchronisation loss refines semantic and beat alignment. Evaluated on AIST++ and PopDanceSet under identical protocols, PF–MambaDance sets new state-of-the-art rhythm accuracy (AIST++ BHR/BC 98.9/96.8; PopDanceSet 97.6/95.1) and improves physical plausibility (MPJPE 10.9 mm, foot-slide 0.91 %, bone-error 0.65 % on AIST++), as confirmed by cross-genre qualitative results. The integration of state-space temporal memory, graph-aware kinematics and anatomical priors within diffusion closes key gaps in beat precision, phrase coherence and biomechanical realism for practical music-conditioned motion synthesis.</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

Pf-mambadance: pose-fusion prior guided Mamba-diffusion for music-driven 3D dance generation

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Xinqiao Liu,
  • Anming Dong,
  • Rui Zhang

摘要

Long-horizon alignment of musical structure with biomechanically valid motion remains elusive for 3D dance generation: existing diffusion models drift across phrases and produce foot-skate artifacts. We introduce PF–MambaDance, a conditional diffusion framework that factorises temporal and spatial modelling. A linear-time Mamba state-space backbone captures phrase-level dynamics, an anatomical graph-CNN enforces skeletal coupling, and multi-scale FiLM fuses beat/phrase cues. A Pose-Fusion prior (2D reprojection cycle with a frozen OpenPose-CNN-GCN) and lightweight physics regularisers jointly constrain bone lengths, joint limits and ground contacts, while a music \(\leftrightarrow \) dance cycle head with lag-robust synchronisation loss refines semantic and beat alignment. Evaluated on AIST++ and PopDanceSet under identical protocols, PF–MambaDance sets new state-of-the-art rhythm accuracy (AIST++ BHR/BC 98.9/96.8; PopDanceSet 97.6/95.1) and improves physical plausibility (MPJPE 10.9 mm, foot-slide 0.91 %, bone-error 0.65 % on AIST++), as confirmed by cross-genre qualitative results. The integration of state-space temporal memory, graph-aware kinematics and anatomical priors within diffusion closes key gaps in beat precision, phrase coherence and biomechanical realism for practical music-conditioned motion synthesis.