Motion Beyond Morphology: Bootstrapping Cross-Category Motion Transfer from Abstract Motion Representations
The paper targets motion transfer when the source and target do not share the same body structure.
The authors propose a two-stage framework that learns abstract motion views, then uses them to create cross-category video pairs for training. A second stage folds that supervision into direct reference-video-conditioned generation, so inference does not require explicit motion extraction. They also introduce OpenVMT-Dataset and OpenVMT-Bench to test transfer across same, near, and far category gaps. Experiments are reported as state of the art for motion fidelity and target preservation. HF Daily Papers' note
The authors propose a two-stage framework that learns abstract motion views, then uses them to create cross-category video pairs for training. A second stage folds that supervision into direct reference-video-conditioned generation, so inference does not require explicit motion extraction. They also introduce OpenVMT-Dataset and OpenVMT-Bench to test transfer across same, near, and far category gaps. Experiments are reported as state of the art for motion fidelity and target preservation. HF Daily Papers' note
score 4