Identity-aware motion modeling
Texture-rich image latents are replicated and warped along joint trajectories, replacing ambiguous sparse-to-dense attention alignment.
IAMMResearch project · 2026
Let the latent
dance.
Identity-aware motion modeling for realistic, expressive character animation.
Self-driven animation
The idea
Motion should move with identity.
Most image-to-video systems inject sparse pose signals and ask attention to align them with dense character pixels. LatentDance takes a different route: it warps the character's own latent features under pose guidance, creating one unified, identity-aware motion representation.
01 / Method
Three pieces,
one latent space.
A compact design for motion that feels natural and characters that stay themselves.
Pseudo-latent completion
Joint-aware semantic embeddings fill missing or occluded regions, letting the generative prior synthesize a complete structure.
PLCLatent identity preservation
A direct cross-attention shortcut transfers clean first-frame details to generated frames and prevents identity drift.
LIP02 / Results
Natural motion.
Consistent characters.
Across real people, stylized characters, and fast motion, the latent stays grounded.
19.194PSNR ↑
0.8281SSIM ↑
12.265FID-VID ↓
228.159FVD ↓
Takeaway
Better motion starts
with a clean latent.
LatentDance establishes a simpler path to animation: preserve what makes a character recognizable, then let motion unfold around it.
Get the code ↗Citation
LatentDance
@article{yang2026latentdance,
title={LatentDance: Towards Realistic and Dynamic Character Animation via Identity-Aware Motion Representation},
author={Yang, Yixin and Jin, Yeying and Zhang, Jiawei and Sun, Long and Cheng, Xu and Pan, Jinshan},
year={2026}
}