Reproduce “STEER: Steerable Dyadic Head Avatars” 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
Kartik Teotia, Helge Rhodin, Hyeongwoo Kim, Marc Habermann, Christian Theobalt
Researchers at the Max Planck Institute for Informatics developed STEER, a system that generates steerable, dyadic head avatar motion with explicit control over gaze, head rhythm, and emotion. This approach combines partner-aware motion generation with photorealistic rendering, demonstrating superior realism, diversity, and reactivity in quantitative and qualitative evaluations.
Key insights
- Detailed non-verbal behavioral pseudo-labels can be effectively extracted from unconstrained dyadic video, providing essential supervision for controllable avatar models.
- A two-phase training and distillation strategy allows semantic control pathways to be injected without degrading the core audio-and-partner motion generation backbone.
- Disentangled, real-time control over non-verbal cues can be achieved in photorealistic avatars by mapping generated parametric motion to a frozen Gaussian avatar prior and independently modulating gaze.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.