Research map / Video / Video segmentation and tracking
MoST: Motion Style Transformer Between Diverse Action Contents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MoST, a transformer-based motion style transfer model with Siamese content/style encoders, a part-attentive style modulator, and a style disentanglement loss, enabling style transfer between motions with different content without heuristic post-processing.
Problem
Motion style transfer fails when the style motion and content motion have different action contents, because style and content are not cleanly separated; prior methods also rely on heavy post-processing (copying global translation, foot-contact fixing).
Approach
Siamese transformer encoders extract style and content features from each motion with body-part attention; a part-attentive style modulator cross-attends style features to the target content so style is transferred to the right body parts; a style disentanglement loss forces style features to be invariant to the content of the style motion.
Key results
- On Xia dataset, MoST achieves average CC 8.5 and SC++ 63.0, the lowest of all methods (best baseline Wen et al. CC 18.5, Park et al. SC++ 65.7).
- With different-content pairs, MoST gets CC 8.7 and SC++ 63.2, while MotionPuzzle reaches SC++ 82.5 and Aberman et al. CC 47.0.
- Ablation: adding PSM and style disentanglement loss together reduces CC from 37.4 to 8.5 and SC++ from 69.5 to 63.0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Wen et al. with MoST | motion style transfer · Xia dataset, different content | CC: 22.3 -> 8.7 | Helped |
| Replace Park et al. with MoST | motion style transfer · Xia dataset, different content | SC++: 67.9 -> 63.2 | Helped |
| Add PSM (ablation) | MoST without style disentanglement loss · Xia dataset | CC: 37.4 -> 19.7 | Helped |
| Add PSM (ablation) | MoST without style disentanglement loss · Xia dataset | SC++: 69.5 -> 66.1 | Helped |
| Add style disentanglement loss (ablation) | MoST without PSM · Xia dataset | CC: 37.4 -> 9.3 | Helped |
| Add style disentanglement loss (ablation) | MoST without PSM · Xia dataset | SC++: 69.5 -> 63.2 | Helped |
| Add PSM (ablation) | MoST with style disentanglement loss · Xia dataset | CC: 9.3 -> 8.5 | Helped |
| Add PSM (ablation) | MoST with style disentanglement loss · Xia dataset | SC++: 63.2 -> 63.0 | Helped |
Surprising findings
- Existing methods that rely on heuristic post-processing degrade motion quality, and MoST avoids post-processing entirely while still outperforming them.
Limitations
- Some outputs still show foot-contact problems since the physics-based loss does not forcibly remove them.
- Transformer requires a fixed maximum sequence length T (set to 200).
- Evaluated only on two mocap datasets with limited styles and contents.
Experimental setup
Datasets: Xia dataset BFA dataset
Benchmarks: Xia dataset BFA dataset
Metrics: content consistency (CC) style consistency++ (SC++)
Baselines: MotionPuzzle Aberman et al. Park et al. Wen et al.
Builds on: AdaIN StyTr2 MotionPuzzle StarGAN-style domain translation (Park et al.) Wen et al. generative flow motion style transfer
Closest papers by meaning
- Arbitrary Motion Style Transfer with Multi-condition Motion Latent Diffusion Model (CVPR 2024 · similarity 0.85)
- Generative Human Motion Stylization in Latent Space (ICLR 2024 · similarity 0.84)
- Scalable Motion Style Transfer with Constrained Diffusion Generation (AAAI 2024 · similarity 0.84)
- Puff-Net: Efficient Style Transfer with Pure Content and Style Feature Fusion Network (CVPR 2024 · similarity 0.83)
- ConMo: Controllable Motion Disentanglement and Recomposition for Zero-Shot Motion Transfer (CVPR 2025 · similarity 0.83)
- DisMo: Disentangled Motion Representations for Open-World Motion Transfer (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.