Reproduce “Human Motion Diffusion as a Generative Prior” on OpenResearch

About this paper

Yonatan Shafir, Guy Tevet, Roy Kapon, Amit H. Bermano

Researchers at Tel-Aviv University developed techniques to leverage pre-trained motion diffusion models as generative priors, enabling zero-shot generation of arbitrarily long human motion sequences, few-shot text-guided two-person interaction, and fine-grained, composable motion control. This approach successfully extends model capabilities to complex, data-scarce tasks, outperforming prior methods and generating coherent, controlled, and diverse motions.

Key insights

• Pre-trained diffusion models can act as robust generative priors, allowing complex compositional tasks to be addressed with minimal or no additional data.
• Explicitly managing temporal consistency via 'handshakes' during denoising enables zero-shot generation of arbitrarily long sequences from models trained on short clips.
• Learning only the 'communication' between separate person-specific priors, rather than training a full multi-person model, is effective for few-shot multi-person interaction generation.

Summary by alphaXiv.