Research map / Video / Video generation
DeforM: Reasoning-Guided Physics-Aware Video Generation via Spatial-Temporal Masking
Yunyi Li, Yu Qiao, Yaohui Wang, Xinyuan Chen
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 VLM-guided masks (DeforM-Reason) instead of random masks (ablation) | DeforM-Injection · Ablation (Wan2.2-I2V-5B, deformation test set) | Physical Score: 3.82 -> 4.40 | Improved |
| add VLM-guided masks (DeforM-Reason) instead of random masks (ablation) | DeforM-Injection · Ablation (Wan2.2-I2V-5B, deformation test set) | Dynamic Degree: 47.58 -> 54.74 | Improved |
| add training-based mask injection over the training-free variant (ablation) | DeforM-Free · Ablation (Wan2.2-I2V-5B, deformation test set) | Physical Score: 3.89 -> 4.40 | Improved |
| add training-free mask guidance to vanilla Wan2.2 (ablation) | Wan2.2-I2V-5B · Ablation (Wan2.2-I2V-5B, deformation test set) | Physical Score: 3.49 -> 3.89 | Improved |
| training-free DeforM-Free versus standard SFT (ablation) | Wan2.2-I2V-5B · Ablation (Wan2.2-I2V-5B, deformation test set) | Physical Score: 3.76 -> 3.89 | Improved |
| standard SFT on deformation data versus vanilla (ablation) | Wan2.2-I2V-5B · Ablation (Wan2.2-I2V-5B, deformation test set) | Dynamic Degree: 42.61 -> 44.77 | No clear effect |
| add self-attention injection to cross-attention modulation (ablation) | DeforM-Cross · Ablation (Wan2.2-I2V-5B, deformation test set) | Physical Score: 4.09 -> 4.40 | Improved |
| add cross-attention modulation to self-attention injection (ablation) | DeforM-Self · Ablation (Wan2.2-I2V-5B, deformation test set) | LDF: 3.33 -> 4.16 | Improved |
| self-attention injection instead of cross-attention only (ablation) | DeforM-Cross · Ablation (Wan2.2-I2V-5B, deformation test set) | Motion Smoothness: 98.43 -> 98.62 | Mixed |
| DeforM-Injection versus Wan2.2-I2V-14B baseline | Wan2.2-I2V-14B · Deformation test set (320 samples) | Physical Score: 3.95 -> 4.40 | Improved |
| DeforM-Injection versus Hunyuan-I2V | Hunyuan-I2V · Deformation test set (320 samples) | Dynamic Degree: 27.74 -> 54.74 | Improved |
| DeforM-Injection versus Hunyuan-I2V temporal flickering | Hunyuan-I2V · Deformation test set (320 samples) | Temporal Flickering: 99.30 -> 97.24 | Hurt |
Closest papers by meaning
- PhysVid: Physics Aware Local Conditioning for Generative Video Models (CVPR 2026 · similarity 0.89)
- PhysAlign: Learning Physical Priors for Dynamical Event-Driven Video Generation via Representation Alignment (ECCV 2026 · similarity 0.88)
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · similarity 0.88)
- PhysChoreo: Physics-Controllable Video Generation with Part-Aware Semantic Grounding (ECCV 2026 · similarity 0.88)
- DeformMaster: An Interactive Physics-Neural World Model for Deformable Objects from Videos (NeurIPS 2026 · similarity 0.87)
- Physics-Informed Video Generation via Mixture-of-Experts Latent Alignment (NeurIPS 2026 · 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.
