Research map / Video / Video generation
Controlling Motion Transfer in Diffusion Transformers via Attention Heads
Sunyoung Jung, Jiwoo Park, Yoonseok Choi, Kyobin Choo, Ming-Hsuan Yang, Seong Jae Hwang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add selective structural head injection (ablation) | HALO · Motion transfer benchmark | Motion Fidelity: 61.3 -> 66.2 | Improved |
| add semantic-aware motion guidance (ablation) | HALO · Motion transfer benchmark | Motion Fidelity: 61.8 -> 66.2 | Improved |
| add both components to displacement optimization (ablation) | HALO · Motion transfer benchmark | Motion Fidelity: 59.6 -> 66.2 | Improved |
| add both components to displacement optimization (ablation) | HALO · Motion transfer benchmark | Temporal Consistency: 89.5 -> 87.5 | Hurt |
| use motion-specific heads instead of all heads (ablation) | HALO · Motion transfer benchmark | Motion Fidelity: 63.1 -> 66.2 | Improved |
| use motion-specific heads instead of spatial heads (ablation) | HALO · Motion transfer benchmark | Motion Fidelity: 53.6 -> 66.2 | Improved |
| inject low-entropy heads (tau<7) rather than high-entropy (ablation) | HALO · Motion transfer benchmark | Motion Fidelity: 54.1 -> 66.2 | Improved |
| use top-k 4 instead of 10 (ablation) | HALO · Motion transfer benchmark | Motion Fidelity: 45.2 -> 66.2 | Improved |
| use semantic bias beta 0.1 instead of 0.3 (ablation) | HALO · Motion transfer benchmark | Motion Fidelity: 53.6 -> 66.2 | Improved |
| use 12 rather than 10 optimization steps (ablation) | HALO · Motion transfer benchmark | Motion Fidelity: 50.1 -> 66.2 | Improved |
| HALO vs GWTF on Movie Scene Dataset | HALO · Movie Scene Dataset | Motion Fidelity: 46.6 -> 52.6 | Improved |
| HALO vs GWTF on hard motion | HALO · Motion transfer benchmark (Hard motion) | Motion Fidelity: 56.0 -> 59.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- RoPECraft — RoPECraft: Training-Free Motion Transfer with Trajectory-Guided RoPE Optimization on Diffusion Transformers (NeurIPS 2025) · better on 4 of 4 shared results
- MotionClone — MotionClone: Training-Free Motion Cloning for Controllable Video Generation (ICLR 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Video Motion Transfer with Diffusion Transformers (CVPR 2025 · similarity 0.92)
- DiTCtrl: Exploring Attention Control in Multi-Modal Diffusion Transformer for Tuning-Free Multi-Prompt Longer Video Generation (CVPR 2025 · similarity 0.89)
- Emergent Temporal Correspondences from Video Diffusion Transformers (NeurIPS 2025 · similarity 0.88)
- Video Diffusion Models are Training-free Motion Interpreter and Controller (NeurIPS 2024 · similarity 0.88)
- VD3D: Taming Large Video Diffusion Transformers for 3D Camera Control (ICLR 2025 · similarity 0.87)
- Space-Time Diffusion Features for Zero-Shot Text-Driven Motion Transfer (CVPR 2024 · similarity 0.87)
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.
