Motion4Motion offers training-free cross-species motion transfer

A new arXiv preprint from Ling-Hao Chen and co-authors, Motion4Motion, takes a swing at one of the older pain points in generative animation: video-to-video motion transfer only really works when both characters share a human-like skeleton, because that is what the models were trained to condition on.

The argument the authors make is straightforward. Skeleton-conditional pipelines rely on a predefined human skeleton structure and require skeleton-conditional model training, which the paper says makes them difficult to generalize to diverse characters, such as animals from different species, while preserving their unique motion styles. They also point out that labeled data in diverse skeletons is limited, which restricts the large-scale training that would otherwise fix the generalization problem. In other words, the bottleneck is not the model, it is the annotation stack behind it.

Motion4Motion's move is to drop the skeleton entirely. Instead of predicting joints and re-targeting them onto a new rig, the framework models the motion flow of the character in a video, which the authors argue makes motion transfer across species easier. They describe it as training-free, meaning the transfer is done at inference rather than by fitting a new model per character class, and they claim their method outperforms baselines impressively on the applications shown on the project page.

The honest caveat is that the abstract is where the specificity stops. It does not name the baselines, does not give quantitative benchmark numbers, and does not describe how far from humanoid the target characters can go before the flow-based approach breaks. Take the claim as the authors' framing, not a settled result, until the full paper and the project-page examples are examined side by side.

If it does hold up under scrutiny, the interesting downstream effect is on the studios and toolmakers whose current advantage is a carefully labeled skeleton library for creatures. A training-free primitive that works across species is exactly the kind of thing that quietly rearranges who has to build what in an animation pipeline.