Pipeline

X-UniMotion features an end-to-end training framework that jointly learns an implicit latent representation of full-body human motion and synthesizes lifelike videos using a DiT network. At its core, we employ an image encoder $\mathcal{E}$ to extract a 1D latent motion descriptor $z$ from the driving image $I_D$, capturing full-body articulation. This global motion code is complemented by decoupled local descriptors—$z_{lh}$ and $z_{rh}$ for the left and right hands, and $z_f$ for facial expressions—extracted from local patches corresponding to the hands and face via $\mathcal{E}_h$ and $\mathcal{E}_f$ respectively. To achieve identity-agnostic motion representation, we apply spatial and color augmentations that effectively disentangle identity cues from the motion latents. These motion tokens are retargeted to the reference subject's body structure in $I_S$ via a ViT decoder $\mathcal{D}$, which outputs identity-aligned spatial motion guidance. This guidance is concatenated with noised video latents and, together with the reference image latents, provided as input to the DiT model. The facial motion latent $z_f$ is further injected into the DiT network via cross-attention layers to control expressions. To supervise motion encoding, we apply dual decoding $\mathcal{D}_h$ and $\mathcal{D}_n$ on the intermediate motion features, predicting both joint heatmaps and hand normal maps. During inference, latent motion codes are directly extracted from each frame of the driving video, enabling expressive and photorealistic animations that maintain strong identity resemblance to the reference image.

Results

Diversity