IM-Animation | CVPR 2026 Findings Track

📄 Abstract

Recent progress in video diffusion models has advanced character animation. Explicit methods represent motion using skeleton or DWPose, but struggle with spatial mismatches. Implicit methods capture motion semantics directly, but suffer from identity leakage.

We propose a novel implicit motion representation compressing per-frame motion into 1D motion tokens. We design a mask token-based retargeting module with temporal training bottleneck. Our three-stage training strategy ensures high fidelity.

IM-Animation achieves superior performance vs. state-of-the-art.

🎯 Teaser

IM-Animation introduces an impressive implicit motion representation and retargeting method. Our model supports implicit video model motion control in cases with significant scale differences or substantial variations in posture and body shape.

💡 Key Contributions

We propose IM-Animation, an implicit portrait animation solution. Given a identity image and a motion video, we employ a three-stage training strategy. In the first stage, we train a compact motion encoder based on a 1D tokenizer. In the subsequent second and third stages, we train a temporal retargeting module based on mask tokens, utilizing a lightweight heatmap decoder for intermediate supervision. This approach ensures that we can encode precise retargeted information without disclosing the ID information of the driven video or the pose information of the source image. Ultimately, we achieve end-to-end training of the entire model.

1

Compact 1D Motion Representation

Compresses motion into spatially invariant 1D tokens, preventing identity leakage.

2

Mask Token-based Retargeting

Uses mask tokens as latent bottleneck to remove motion from source image.

3

Efficient Three-Stage Training

Progressive training reduces costs and achieves effective disentanglement.

▶️ Qualitative Results

📊 Quantitative Results (Cross Reenactment)

Bold = Best, Highlighted = Second Best

Method	PSNR ↑	SSIM ↑	LPIPS ↓	FID ↓	FVD ↓
Champ [78]	20.42	0.89	0.28	88.09	785.54
MimicMotion [69]	20.01	0.88	0.30	97.86	901.55
AnimateAnyone [22]	19.24	0.86	0.31	85.24	690.15
UniAnimate-DiT [54]	21.97	0.90	0.25	71.61	436.80
Animate-X [45]	22.16	0.90	0.27	75.38	624.24
Wan-Animate [9]	22.72	0.92	0.24	59.10	267.71
IM-Animation (Ours)	22.87	0.91	0.24	51.19	270.42

📚 Citation

@inproceedings{xu2025imanimation,
  title={IM-Animation: An Implicit Motion Representation 
         for Identity-decoupled Character Animation},
  author={Xu, Zhufeng and Gao, Xuan and Liu, Feng-Lin and 
          Zhang, Haoxian and Fang, Zhixue and Lai, Yu-Kun and 
          Liu, Xiaoqiang and Wan, Pengfei and Gao, Lin},
  booktitle={CVPR},
  year={2025}
}