OmniHumanoid: Streaming Cross-Embodiment Video Generation with Paired-Free Adaptation

Cross-embodiment video generation aims to transfer motions across different
humanoid embodiments, such as human-to-robot and robot-to-robot, enabling scalable data generation for embodied intelligence. A major challenge in this setting
is that motion dynamics are partly transferable across embodiments, whereas appearance and morphology remain embodiment-specific. Existing approaches often
entangle these factors, and many require paired data for every target embodiment,
which limits scalability to new robots. We present OmniHumanoid, a framework
that factorizes transferable motion learning and embodiment-specific adaptation.
Our method learns a shared motion transfer model from motion-aligned paired
videos spanning multiple embodiments, while adapting to a new embodiment using
only unpaired videos through lightweight embodiment-specific adapters. To reduce interference between motion transfer and embodiment adaptation, we further
introduce a branch-isolated attention design that separates motion conditioning
from embodiment-specific modulation. In addition, we construct a synthetic crossembodiment dataset with motion-aligned paired videos rendered across diverse
humanoid assets, scenes, and viewpoints. Experiments on both synthetic and realworld benchmarks show that OmniHumanoid achieves strong motion fidelity and
embodiment consistency, while enabling scalable adaptation to unseen humanoid
embodiments without retraining the shared motion model. Code is released at
https://github.com/showlab/OmniHumanoid