局部生成和全局融合的文本驱动三维人体动作扩散生成

Key Points

• This research aims to improve the process of generating 3D human movements from textual prompts by addressing existing limitations.
• Proposed a cascading diffusion generation framework integrating local and global features.
• Utilized a large language model to independently describe body parts for action generation.
• Developed a global fusion network to combine local features into biomechanically accurate full-body postures.
• Demonstrated improvements in generation quality and action accuracy compared to baseline methods, with detailed evaluations on HumanML3D and KIT-ML datasets.
• Achieved higher scores in FID and CLIP-S metrics compared to baseline methods.