Motion4Motion

Motion4Motion is a research framework for transferring motion from a source video to a target subject, including across species with different body structures. Instead of relying on a predefined skeleton, it models the subject's motion as a dense motion flow of tracked anchor points, and injects that flow into a pre-trained video diffusion transformer (WAN-T2V) at inference time. No additional training is required.\n\nThe pipeline has two stages. First, anchor points are sampled on the source subject with Grounded SAM-2, matched to the target using a diffusion-feature semantic matcher, and tracked over time with a point tracker. Second, the method, called TransPE (Transferring Positional Encoding), caches the target subject's keys/values, replicates them along time, and re-embeds them with RoPE positions taken from the retargeted motion flow, so the denoiser places the target's appearance along the source trajectories.\n\nThe authors report evaluations on animal (33 pairs) and human (123 pairs) benchmarks, claiming better results than FlexiAct, MotionClone, MotionDirector, RoPECraft, Diffusion-As-Shader and WAN-Move across text similarity, motion fidelity, temporal consistency, appearance consistency and pose similarity. A demonstration drives a table's walking gait from a human video using SAM-2 masks to bind human legs to table legs. The work is a SIGGRAPH 2026 conference paper. At the time of the provided material, no code, weights or demo had been linked; only the project page was available.