Research map / Video / Video generation
PhyMAGIC: Physical Motion-Aware Generative Inference with Confidence-guided VLM
Siwei Meng, Yawei Luo, Ping Liu
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 iterative VLM-guided motion probing with confidence gating over one-shot VLM (ablation) | PhyMAGIC · five-scene subset | material accuracy (%): 60.0 -> 90.0 | Improved |
| add confidence gating to iterative probing (ablation) | PhyMAGIC · five-scene subset | material accuracy (%) and probes/scene: 83.0 -> 90.0 accuracy; 1.00 -> 0.80 probes/scene | Improved |
| add best motion template over fixed boundary for one-shot VLM (ablation) | One-shot VLM · five-scene subset | material accuracy (%): 60.0 -> 76.0 | Improved |
| increase refinement iterations from 1 to 3 (ablation) | PhyMAGIC · Driving car | accuracy (%): 29.17 -> 93.75 | Improved |
| increase refinement iterations from 1 to 3 (ablation) | PhyMAGIC · Swing ficus | accuracy (%): 62.92 -> 90.63 | Improved |
| increase refinement iterations from 1 to 3 (ablation) | PhyMAGIC · Rolling basketball | accuracy (%): 50.27 -> 90.00 | Improved |
| add confidence-driven prompt refinement to CogVideoX (ablation) | CogVideoX · human evaluation plausibility | physical plausibility: 2.58 -> 2.97 | Improved |
| add confidence-driven prompt refinement to CogVideoX (ablation) | CogVideoX · CLIP similarity (8 scenarios, average) | CLIP similarity: 0.239 -> 0.240 | No clear effect |
| full PhyMAGIC pipeline (physics simulation) vs CogVideoX* on aesthetic score | PhyMAGIC · Aesthetic score (8 scenarios, average) | aesthetic score: 31.86 -> 30.69 | No clear effect |
| replace Physics3D with PhyMAGIC for physics-aware generation | PhyMAGIC · CLIP similarity (6 scenes) | CLIP similarity: 0.217 -> 0.252 | Improved |
| PhyMAGIC vs Physics3D on image-motion distribution | PhyMAGIC · Image-Motion-FID | Image-Motion-FID: 98.91 -> 94.69 | Improved |
| use OpenSora-2 instead of CogVideoX-5B as I2V generator (ablation) | PhyMAGIC · inference accuracy gain | accuracy gain (%): +12.2 -> +10.9 | No clear effect |
Closest papers by meaning
- PhyCo: Learning Controllable Physical Priors for Generative Motion (CVPR 2026 · similarity 0.92)
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · similarity 0.90)
- PhysVid: Physics Aware Local Conditioning for Generative Video Models (CVPR 2026 · similarity 0.90)
- Dynamics: Language-Based Representation for Inferring Rigid-Body Dynamics From Videos (CVPR 2026 · similarity 0.90)
- PHANTOM: Physics-Infused Video Generation via Joint Modeling of Visual and Latent Physical Dynamics (CVPR 2026 · similarity 0.89)
- Physics-Informed Video Generation via Mixture-of-Experts Latent Alignment (NeurIPS 2026 · 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.
