Two2Four, a Generative Motion Technology for Controlling Quadrupeds with Human Movement
Two2Four turns ordinary human motion into plausible quadruped animation while preserving intuitive control over actions, head direction, and individual limbs.
Paper Two2Four: Generative Quadruped Puppeteering from Human Motion
Research DisneyResearch|Studios · ETH Zurich
Released 2026-07-30
Converting Ordinary Human Motion into Quadruped Movement
Two2Four is a study that takes ordinary movements performed by humans as input and generates natural, controllable quadruped animations. Previously, animators had to either motion-capture specialized performers who could precisely imitate animal behavior or retarget human movements using complex control systems. Both approaches are difficult to produce and struggle to reproduce the subtle movements of real animals.
The researchers used a two-stage generative diffusion model trained solely on quadruped motion data. Through structured conditional inputs and an inpainting strategy, it transforms human intent into movements suited to an animal’s body structure and handles a range of behaviors, including walking, running, jumping, sitting, and lying down.
In addition to automatically converting an entire motion, the system can control the direction of the head or individual legs separately. This allows animators to use intuitive human movements as the broad outline of a performance while fine-tuning only the body parts that need adjustment.
Experiments reported improved motion realism and controllability over existing retargeting methods. Fatemeh Zargarbashi, Zehong Qiu, Dhruv Agrawal, Stelian Coros, Robert W. Sumner, Martin Guay, and Jakob Buhmann contributed to the work, which is aimed at virtual production and character animation tools.