A new research paper, Motion Beyond Morphology, tackles the challenge of video motion transfer when the reference and target objects differ significantly in shape, articulation, or how they deform. Traditional methods rely on fixed structural correspondence, which fails in such cross-category scenarios.
The proposed framework, MBM, operates in two stages. First, it learns complementary multi-granularity abstract motion representations to bootstrap cross-category video pairs that preserve transferable dynamics. Second, it internalizes this supervision into direct reference-video-conditioned generation, eliminating the need for explicit motion extraction at inference time.
The authors introduce OpenVMT-Dataset and OpenVMT-Bench to train and evaluate image- and text-conditioned motion transfer across three category gaps: Same, Near, and Far. Extensive experiments show that MBM achieves state-of-the-art motion fidelity and target preservation.
The project page and code are publicly available, and the dataset and benchmark are planned for release upon acceptance.