Research map / Video / Video generation
Seeing the Unseen: Unified Visible–Invisible Motion for Physically Consistent Video Generation
Mingyang Bao, Fangda Ye, Xihao Chen, Bobo Li, Xiangtai Li, Shengqiong Wu
Read the paperCheck your idea against 137,217 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 a unified visible-invisible motion framework to inject physical priors into video generation for more physically consistent dynamics (abstract only; details unavailable).
Problem
Video generation models often produce physically inconsistent dynamics because they lack explicit modeling of underlying physical processes; existing simulator-based or coarse semantic guidance approaches have strong assumptions or fail to inject physical priors effectively.
Approach
Unifies visible and invisible motion as a shared representation to inject physical priors into video generation, in contrast to simulator-dependent methods or coarse semantic guidance. Specific mechanism not stated in the available abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add PGDS-simulated visible+invisible motion field to the video backbone | Wan2.2-TI2V-5B · Domino | SA: 2.650 -> 3.976 | Improved |
| add PGDS-simulated motion field to the video backbone | Wan2.2-TI2V-5B · Free Fall | PC: 3.083 -> 4.328 | Improved |
| add future-emergent trajectories to visible-only guidance (ablation) | Ours · Domino | SA: 2.693 -> 3.976 | Improved |
| add future-emergent trajectories to visible-only guidance (ablation) | Ours · VideoPhy2 subset | SA: 3.47 -> 4.01 | Improved |
| add future-emergent trajectories to visible-only guidance (ablation) | Ours · Translation | SA: 3.837 -> 4.013 | Improved |
| use full dual-stream guidance instead of Wan-Move sparse tracks | Ours · Translation | SA: 4.167 -> 4.013 | Hurt |
| use full dual-stream tracking instead of optical flow (Go with Flow) | Ours · Domino | SA: 2.583 -> 3.976 | Improved |
| drop about 25% of object tracks (ablation) | Ours · Robustness avg of 7 scenarios | SA: 4.095 -> 3.890 | Hurt |
| remove 1-2 segmented objects (ablation) | Ours · Robustness avg of 7 scenarios | SA: 4.095 -> 3.733 | Hurt |
| apply both perturbations (ablation) | Ours · Robustness avg of 7 scenarios | SA: 4.095 -> 3.598 | Hurt |
| add dual-stream conditioning (48 to 144 latent channels) | Wan2.2-TI2V-5B · runtime | total generation time: 144.20 s -> 150.28 s | Hurt |
| use LLM-based refinement instead of trajectory guidance (PhysT2V) | PhysT2V · Free Fall | SA vs PC: SA 2.050, PC 4.283 | Mixed |
Limitations
- Only the abstract is available; method details and results are unknown
Closest papers by meaning
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · similarity 0.91)
- How Far Is Video Generation from World Model: A Physical Law Perspective (ICML 2025 · similarity 0.89)
- VideoJAM: Joint Appearance-Motion Representations for Enhanced Motion Generation in Video Models (ICML 2025 · similarity 0.89)
- LaMo: Self-Supervised Latent Motion Priors for Physical Realism in Video Generation (NeurIPS 2026 · similarity 0.88)
- Physics-Informed Video Generation via Mixture-of-Experts Latent Alignment (NeurIPS 2026 · similarity 0.88)
- Towards Physical Understanding in Video Generation: A 3D Point Regularization Approach (NeurIPS 2025 · similarity 0.88)
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.
