Two2Four: Generative Quadruped Puppeteering from Human Motion

Two2Four tackles a long-standing animation problem: getting believable animal motion without relying on expert mocap performers or heavily hand-tuned retargeting setups. The system converts human motion into quadruped behavior using a two-stage generative diffusion pipeline, with training done entirely on quadruped motion data rather than paired human/animal examples.

The practical hook for game teams is controllability. Beyond broad action transfer, the framework supports structured conditioning and inpainting so it can generate walking, running, jumping, sitting, and lying motions while still allowing direct control over details like head movement and individual limbs. That makes it more than a novelty retargeter; it’s aimed at animation workflows where directors and animators need quick iteration without losing intent.

For developers working on creatures, virtual production, or cinematic tools, the interesting part is the shift away from bespoke control rigs toward data-driven motion synthesis. The reported results show better realism and controllability than existing retargeting approaches, which suggests a path toward faster prototyping for animal characters and fewer manual fixes in the animation pass.

The exact production integration details haven’t been disclosed, but the technique points toward a future where human performance can serve as a higher-level input for non-human characters. If it holds up outside the paper setting, it could be useful anywhere teams need believable quadruped motion without building a custom system from scratch.