Research map / Video / Video generation
What Moves? Localized Motion Representations for Compositional Scene Control
Frank Fundel, Malek Ben Alaya, Thomas Ressler-Antal, Stefan Andreas Baumann, Bjorn Ommer
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 |
|---|---|---|---|
| keep full-scene context instead of cropping the RGB input (ablation) | Ours · A2D localized action classification | accuracy: 27.98 -> 33.90 | Improved |
| use actor-specific mask instead of global full-frame mask (ablation) | Ours · A2D localized action classification | accuracy: 21.75 -> 33.90 | Improved |
| use localized motion encoder instead of DisMo | Ours · A2D localized action classification | accuracy: 27.24 -> 33.90 | Improved |
| use localized representation instead of DisMo for motion transfer | CogVideoX-5B · Controllable motion transfer | motion selectivity: 0.1377 -> 0.2304 | Improved |
| use localized representation instead of DisMo for motion transfer | CogVideoX-5B · Controllable motion transfer | out-of-region motion leakage: 0.5604 -> 0.5158 | Improved |
| use localized representation instead of ATI | Wan 2.1-14B vs CogVideoX-5B · Controllable motion transfer | in-region motion fidelity: 0.8111 -> 0.7462 | Hurt |
| use localized representation instead of ATI | Wan 2.1-14B vs CogVideoX-5B · Composable motion transfer | motion selectivity: 0.3291 -> 0.4114 | Improved |
| use localized representation instead of ATI | Wan 2.1-14B vs CogVideoX-5B · Composable motion transfer | in-region motion fidelity: 0.7148 -> 0.6718 | Hurt |
| use localized representation instead of DisMo | global setting · SSv2 zero-shot classification | top-1 accuracy: 24.6 -> 26.5 | Improved |
| use localized representation instead of DisMo | global setting · Diving48 zero-shot classification | top-1 accuracy: 20.9 -> 17.1 | Hurt |
| use localized representation instead of DisMo | SemanticMoments synthetic retrieval · motion retrieval | top-1 accuracy: 90.0 vs 89.6 | No clear effect |
| use localized representation instead of global DisMo | LAOM DCS action probing · mean NMSE | NMSE: 0.6415 -> 0.4339 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- Tora — Tora: Trajectory-oriented Diffusion Transformer for Video Generation (CVPR 2025) · better on 4 of 5 shared results
- DisMo — DisMo: Disentangled Motion Representations for Open-World Motion Transfer (NeurIPS 2025) · better on 5 of 5 shared results
- WanMove — Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance (NeurIPS 2025) · better on 4 of 4 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
- Through-The-Mask: Mask-based Motion Trajectories for Image-to-Video Generation (CVPR 2025 · similarity 0.86)
- DisMo: Disentangled Motion Representations for Open-World Motion Transfer (NeurIPS 2025 · similarity 0.86)
- MotionPro: A Precise Motion Controller for Image-to-Video Generation (CVPR 2025 · similarity 0.86)
- Motion Prompting: Controlling Video Generation with Motion Trajectories (CVPR 2025 · similarity 0.85)
- DynaTokens: Teaching Dynamics to Camera-Controlled Video Models at Test Time (NeurIPS 2026 · similarity 0.85)
- Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance (NeurIPS 2025 · similarity 0.85)
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.
