Generative video models achieve high visual fidelity but often violate basic physical principles, limiting reliability in real-world settings. Prior attempts to inject physics rely on conditioning: frame-level signals are domain-specific and short-horizon, while global text prompts are coarse and noisy, missing fine-grained dynamics. We present PhysVid, a physics-aware local conditioning scheme that operates over temporally contiguous chunks of frames. Each chunk is annotated with physics-grounded descriptions of states, interactions, and constraints, which are fused with the global prompt via chunk-aware cross-attention during training. At inference, we introduce negative physics prompts (descriptions of locally relevant law violations) to steer generation away from implausible trajectories. On VideoPhy, PhysVid improves physical commonsense scores by ≈33% over baseline video generators, and by up to ≈8% on VideoPhy2. These results show that local, physics-aware guidance substantially increases physical plausibility in generative video and marks a step toward physics-grounded video models.
Figures & tables
Figure 2: Procedure for generating physics-grounded local prompts during data annotation.
Figure 3: Architecture of PhysVid showing local information pathways with chunk aware cross-attention. Commonly applied procedures such as tokenization, latent encoding, and decoding are implicit and not shown.
Figure 4: Generation of local and counterfactual prompts during inference.
Method
Params (B)
VideoPhy
VideoPhy2
SA
PC
SA
PC
Wan-1.3B
1.3
0.46
0.24
0.28
0.61
Wan-14B
14
0.52
0.24
0.29
0.59
PhysVid
1.7
0.43
0.32
0.28
0.64
Table 1: Results on VideoPhy and VideoPhy2. We report semantic alignment (SA) and physical commonsense (PC; higher is better). PhysVid (1.7B) achieves the best PC on both benchmarks ( ≈33% relative gain on VideoPhy and over 8% on VideoPhy2 vs. Wan-14B).
Figure 5: VideoPhy PC ( PC ) score by category.
Figure 6: VideoPhy2 PC ( PC ) score by category.
Method
SA
PC
Lavie ‡
0.49
0.28
Lavie $
0.49
0.32
VideoCrafter2 †
0.47
0.36
VideoCrafter2 ‡
0.49
0.35
VideoCrafter2 $
0.50
0.30
VideoCrafter2 ∼
0.24
0.15
Table 2: Quantitative comparisons of our approach with previous works evaluated on the VideoPhy benchmark, including both general and physics-aware generative video methods. Symbols indicate results taken from prior work: WISA [ 48 ] ( † ),VideoREPA [ 66 ] ( \$$ ), Hao et al. [20] ( #),PhyT2V[56](\sim),VideoPhy[3](\ddagger).Ourownbaselinesaremarkedwith\ast$ . Best model scores are underlined. For physics-aware methods with multiple entries, we report the best PC score, with parentheses showing relative improvement over the corresponding baseline.
Method
VideoPhy
VideoPhy2
SA
PC
SA
PC
baseline (Wan-1.3B)
0.4570
0.2401
0.2845
0.6144
fine tuning
0.4174
0.2866
0.2765
0.6261
PhysVid w/o counterfactual guidance
0.4355
0.2924
0.2791
0.6334
PhysVid
0.4302
0.3169
0.2775
0.6411
Table 3: Ablations. The highest scores in each metric are highlighted, whereas the lowest scores are underlined.
Figure 7: Qualitative ablation example on a VideoPhy prompt. Best viewed when zoomed in.
Figure 8: VLM instruction to generate the physics caption for a video chunk
Figure 9: VLM instruction to generate the counterfactual physics caption
Figure 10: VLM instruction to generate the local physics captions for all the chunks to be generated at once during inference
Figure 11: A sample from the data with annotations generated by a VLM ( VLM ). The topmost text is the global prompt and the local annotations are listed sequentially alongside a representative frame from each chunk.
Figure 12: A set of physics grounded and physics counterfactual prompts generated during inference. The representative frames from the generated video chunks are shown on the right.
Figure 13: Additional comparisons between PhysVid and Wan-14B . Captions are from VideoPhy.
Figure 14: Supplementary qualitative results from the ablation study. All prompts are from VideoPhy.
Model
LPIPS ↓
FVD ↓
SSIM ↑
PSNR ↑
Wan 1.3B
0.703
417.352
0.217
8.625
Wan (finetuned)
0.671
302.465
0.239
9.379
PhysVid
0.679
318.087
0.240
9.234
Table 4: Additional metrics based on similarity (2048 sample pairs)
Training
Base architecture
Wan-1.3B
Additional parameters (M)
400
Effective batch size
64
Number of steps
3000
Number of epochs
4
Learning rate (Stage 1: 1000 steps, frozen base layers)