Identity-aware motion modeling
Reference latents are replicated and warped along joint and limb trajectories, replacing ambiguous sparse-to-dense alignment with explicit correspondence.
Character animation · 2026
LatentDance
Towards Realistic and Dynamic Character Animation via Identity-Aware Motion Representation
One reference image. Any motion. Identity stays.
Overview
Motion should move with identity.
Existing methods ask attention to align sparse pose signals with dense character pixels. This ill-posed mapping often produces rigid motion and weak texture correspondence.
LatentDance takes a direct route: it warps the reference character’s own latent features under pose guidance, creating a unified representation of identity and motion. The result is more natural movement with stronger visual consistency.
Video presentation
LatentDance in three minutes.
Method, latent-warp visualization, and qualitative results in one short walkthrough.
Method
A unified latent space for identity and motion.
Instead of reconciling two separate conditions, LatentDance makes the reference identity itself move through time.
Pseudo-latent completion
Joint-aware semantic embeddings supply structure for regions that are occluded or missing from the initial frame.
Latent identity preservation
A direct route to the clean first-frame latent restores fine texture throughout denoising and reduces identity drift.
Inside the representation
Latent warp, not pixel warp.
The visualization decodes local latent patches only to show where identity features travel. The actual operation happens in semantic latent space and extends along skeleton connections.
Video results
Dynamic across styles and scenes.
Every clip shows the reference, driving pose, and generated result from left to right.
01Real-world
02Cartoon
03TikTok
More examples are available in the project repository ↗
Evaluation
Strong motion, faithful identity.
LatentDance improves temporal realism while remaining competitive on frame-level fidelity.
- MimicMotion
- Animate-X
- UniAnimate-DiT
- RealisDance-DiT
- Wan-Animate
- One-to-all Animation
- LatentDance (Ours)
- Ground truth
19.194PSNR ↑
0.8281SSIM ↑
0.2414LPIPS ↓
228.159FVD ↓
TikTok benchmark · best FVD among compared methods. See the paper for complete TikTok and Cartoon results.
Open source
Run LatentDance.
Inference code, pretrained checkpoints, pose preprocessing, and a Gradio demo are available now.
Citation
Cite this work.
@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}
}