Anonymous Submission
ECCV 2026 Workshop 3D in the Era of World Models
Side-by-side comparisons of our counterfactual-refined result with the corresponding baselines. Select an input scene and choose a comparison method below.
Ablation of the Counterfactual Physical Reasoning Module. (a) Ours without the Counterfactual Physical Reasoning Module and (b) Ours. Removing the module leads to unstable dynamics, while the full model produces coherent, instruction-aligned physical behavior.
Following the evaluation protocol of PhyMAGIC, we report CLIP Similarity to measure semantic alignment between the generated motion and its text description, Aesthetic Score to assess the visual quality of the generated video, and PhysReal. Higher values indicate better performance for all metrics.
| Method | CLIP Sim. ↑ | Aesthetic ↑ | PhysReal ↑ |
|---|---|---|---|
| RealWonder | 0.269 | 53.60 | 0.624 |
| RealWonder + PhyMAGIC | 0.273 | 54.28 | 0.635 |
| RealWonder + Ours | 0.278 | 55.14 | 0.647 |
Bold: best. Higher is better for all metrics.
We conducted a two-alternative forced-choice (2AFC) user study with 50 participants to compare our method with physics-based video generation baselines. Values indicate the percentage of participants who preferred RealWonder + Ours in terms of physical plausibility and instruction alignment. Higher values are better.
Physics-based video generation combines physics simulators with generative models to produce visually realistic, physically grounded dynamics. However, the resulting motion critically depends on simulator parameters, whose estimation from a single image remains inherently ambiguous. We introduce counterfactual physical reasoning, which reasons about physical parameters through controlled interventions and their resulting dynamics. Across 30 scenes, our module improves instruction alignment, visual quality, and physical realism over both baselines, and receives 80.0–85.1% preference in a 50-participant 2AFC study.
