Reproduce “InterControl: Zero-shot Human Interaction Generation by Controlling Every Joint” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhenzhi Wang, Jingbo Wang, Yixuan Li, Dahua Lin, Bo Dai

A framework called InterControl generates zero-shot multi-person interactions by transforming high-level descriptions into precise spatial joint controls for any group size. It uses a Motion ControlNet and Inverse Kinematics guidance guided by an LLM planner to achieve highly controllable animations with low spatial errors and high user preference compared to prior single-person focused models.

Key insights

  • Multi-person interaction generation can be effectively re-framed as a highly controllable single-person motion synthesis problem, circumventing the need for specialized multi-person training data.
  • The synergy of a conditional generative model (MDM with ControlNet) and a precise real-time optimization method (IK Guidance) enables both plausible motion generation and accurate adherence to fine-grained spatial constraints.
  • Large Language Models provide a practical zero-shot solution for translating abstract natural language interaction descriptions into explicit joint-level control signals, enhancing user accessibility.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.