T2VPhysBench: A First-Principles Benchmark for Physical Consistency in Text-to-Video Generation

• Existing benchmarks commonly use pixel-level metrics and scenario-based prompts, motivating a human-centered, first-principles evaluation.Source: §1 · p. 3
• T2VPhysBench systematically evaluates modern text-to-video models against twelve physical laws spanning Newtonian mechanics, conservation principles, and phenomenological effects.Source: §1 · p. 3
• All evaluated models average below 0.60 in every law category, indicating persistent failure to satisfy basic physical constraints.Source: §1 · p. 3
• Detailed law-specific prompt hints do not overcome models’ inability to generate physically coherent videos.Source: §1 · p. 3
• Counterfactual prompts often induce rule-violating videos, suggesting reliance on surface patterns rather than true physical reasoning.Source: §1 · p. 3

• Existing text-to-video benchmarks cover fidelity, compositionality, temporal dynamics, object counting, and storytelling, but systematic physical-constraint evaluation remains underexplored.Source: §2 · p. 4
• VideoPhy uses human evaluation for material collisions, whereas Physics-IQ evaluates frame extensions with automated metrics such as MSE or IoU.Source: §2 · p. 4
• Prior physics benchmarks generally do not combine first-principles physical laws with careful human evaluation, motivating this benchmark.Source: §2 · p. 4

3.1 Baseline Models

• Ten state-of-the-art models, spanning open- and closed-source systems, are evaluated with short 16:9 videos at typically 720p resolution.Source: §3.1 · p. 5

3.2 Benchmark Prompts

• The benchmark derives twelve laws from Newtonian, conservation, and phenomenological categories, using seven realistic prompts per law for 84 prompts per model.Source: §3.2 · p. 6
• Newtonian prompts test inertia, force–acceleration, action–reaction, and gravitation across free-motion, applied-force, interaction, and orbital scenarios.Source: §3.2 · p. 6
• Conservation prompts examine energy, mass, linear momentum, and angular momentum through transformations, transfers, collisions, and rotating systems.Source: §3.2 · p. 6
• Phenomenological prompts assess Hooke’s, Snell’s, reflection, and Bernoulli principles through elastic, optical, reflective, aerodynamic, and fluid effects.Source: §3.2 · p. 6

3.3 Evaluation Protocol

• Three annotators independently assign four adherence levels mapped to scores from 0.0 to 1.0, which are averaged across prompts and annotators.Source: §3.3 · pp. 6–7

4.1 Overall Physical Constraint Results

• All ten models score below 0.60 on basic physical laws, with Wan 2.1 reaching only 0.56 on Newton’s principles and 0.29 on conservation laws.Source: §4.1 · p. 7, Table 2 · p. 7
• Conservation principles are consistently harder than Newton’s or phenomenon principles across models.LTX Video scores 0.13 on conservation versus 0.40 on both Newton’s and phenomenon principles; Pika 2.2 scores 0.19 versus 0.38 and 0.40.Source: §4.1 · pp. 7–8
• The best overall model, Wan 2.1, averages 0.42 versus 0.19 for SD Video, a 0.23 performance gap.Mochi-1 and Kling achieve only 0.34 and 0.35, respectively.Source: §4.1 · p. 8

4.2 Impact of Hint Levels

• Increasing prompt-hint detail generally does not improve physical-law scores and can sometimes reduce them.Exceptions include Newton’s third law improving from 0.25 to 0.88 and universal gravitation from 0.25 to 0.50; Hooke’s law falls from 0.50 to 0.38, while Newton’s first law falls from 0.12 to 0.00.Source: §4.2 · p. 10

4.3 Impact of Counterfactual Prompts

• Under counterfactual prompts, all models score poorly across physical-law classes, with rare scores above 0.50.Kling scores 0.19 on Newton’s principles and 0.00 on phenomenon principles, while Dreamina scores 0.06 on phenomenon principles.Source: §4.3 · p. 11, Table 3 · p. 11
• Performance under standard prompts does not reliably transfer to counterfactual prompts, indicating that apparent compliance may reflect surface-level pattern matching rather than genuine physical reasoning.Wan 2.1 drops from a standard average of 0.42 to 0.31 counterfactually, while Kling falls from 0.38 to 0.00 on phenomenon principles.Source: §4.3 · p. 11

Understanding and Prediction in World Foundation Models

• World Foundation Models ground video generation in learned physical and spatial constraints, enabling plausible motion and cause-and-effect interactions.They can model acceleration, pushing, and stacking more consistently than systems focused on text–video matching.Source: §5 · p. 12

Rule-based Machine Learning

• Explicit rules, constraints, symbolic reasoning, and physics-informed losses could embed physical laws into training rather than relying only on text–video matching.Differential equations can be incorporated through automatic differentiation, as in physics-informed neural networks.Source: §5 · p. 12
• Hybrid neuro-symbolic systems could combine generative video proposals with symbolic physics engines or differentiable simulators that evaluate and refine physical outcomes.For collisions, a symbolic module could compute the outcome using established equations of motion and enforce it in generated frames.Source: §5 · p. 12

6 Conclusion

• Conservation principles are especially challenging, while Newton’s laws and phenomenological effects show comparatively different performance.The conclusion reports marked variation across law categories, indicating uneven modeling of physical aspects.Source: §6 · p. 12
• All evaluated models struggle with basic Newtonian, conservation, and phenomenological constraints despite impressive visual fidelity and instruction following.The benchmark is human-evaluated and first-principle-inspired.Source: §6 · p. 12
• Detailed prompt hints provide little benefit and can even degrade compliance, indicating that the core limitations extend beyond prompt design.The conclusion describes progressively more detailed hints as ineffective for improving physical consistency.Source: §6 · p. 12
• Counterfactual prompts often produce physically impossible, rule-violating videos, suggesting reliance on pattern memorization rather than true physical reasoning.The finding comes from tests that explicitly request physically impossible scenarios.Source: §6 · p. 12

Appendix

• The appendix contains model details, limitations, societal implications, and detailed video examples in Sections A through D.Source: Appendix · p. 18

A Implementation Details

• Kling supports multiple versions, quality modes, creative parameters, flexible durations and aspect ratios, prompt options, parallel generation, and seed selection.Generation takes approximately four minutes per sample with a batch limit of five videos.Source: A Implementation Details · p. 18
• Wan 2.1 provides Fast and Professional variants, multiple aspect ratios, extended prompts, Inspiration Mode, and generated sound.Source: A Implementation Details · p. 18
• Sora offers multiple resolutions, aspect ratios, durations, style presets, and parallel generation, with access varying by subscription tier.The described subscriptions range from $20 monthly access for 480p and 720p videos to $200 for longer 1080p videos.Source: A Implementation Details · p. 18
• Mochi-1 supports 480p, 16:9 five-second videos at 24FPS, random prompt suggestions, seed control, and two parallel generations.Each generated video takes about three minutes to process.Source: A Implementation Details · p. 18
• LTX Video supports preset styles, several aspect ratios, 512p five-second clips at 24FPS, shot and scene controls, references, voiceover scripts, and staged scene-motion generation.Its workflow first generates the initial scene and then generates motion for that scene.Source: A Implementation Details · pp. 18–19
• Pika 2.2 provides multiple creative features, resolutions, aspect ratios, durations, negative prompts, seeds, and four parallel outputs.Each output takes about 30 seconds according to the implementation description.Source: A Implementation Details · p. 19
• Dreamina includes four variants, prompt enhancement using Deepseek-R1, many aspect ratios, and durations from 3 to 12 seconds at 24FPS.Source: A Implementation Details · p. 19
• The appendix documents Qingying, Hailuo, Stable Video Diffusion, and other evaluated generators across commercial and open-source systems.Qingying is commercial, Hailuo is closed-source, and Stable Video Diffusion is open-source.Source: A Implementation Details · p. 19

• The study documents where and how models fail but does not provide theoretical analyses or guarantees explaining their difficulty with physical constraints.Source: B Limitations · p. 19
• Manual annotation captures nuanced human judgments but limits scalability, making larger model or law coverage costly without reliable automated proxies.Source: B Limitations · p. 19

• The benchmark exposes physics violations relevant to robotics, autonomous vehicles, and scientific visualization, where physical realism supports safety and reliability.Source: C Impact Statement · pp. 19–20
• Researchers and practitioners can use the benchmark to flag unsafe or misleading outputs, guide physics-aware architectures, and curate more robust training datasets.Source: C Impact Statement · p. 20
• The authors foresee no immediate negative societal implications from the benchmark itself, while cautioning that future improvements should avoid unintended cultural or social biases.Source: C Impact Statement · p. 20

• Figures 7–30 present generated video samples for the benchmark prompts, with each figure showing five models and five key frames selected over time.Source: D Video Examples · p. 20
• The examples cover Newtonian mechanics, including Newton’s First, Second, and Third Laws, with a paddling boat illustrating action–reaction behavior.Source: Figure 7 · p. 21, Figure 8 · p. 22, Figure 9 · p. 23
• Additional samples target universal gravitation and conservation laws for energy, mass, momentum, and angular momentum.Source: Figure 12 · p. 26, Figure 13 · p. 27, Figure 14 · p. 28
• The conservation-of-energy example uses a collision in which one ball stops and the other moves, providing a concrete prompt for the displayed samples.Source: Figure 18 · p. 32
• The examples also test phenomenological laws, including Hooke’s Law, Snell’s Law, the Law of Reflection, and Bernoulli’s Principle.Source: Figure 19 · p. 33, Figure 20 · p. 34, Figure 21 · p. 35
• A reflection prompt asks for a diagonally thrown ball to bounce diagonally from a wall, illustrating the kind of concrete scenario used in the samples.Source: Figure 24 · p. 38, Figure 25 · p. 39, Figure 26 · p. 40