PhyGround Framework: 13 Fundamental Physical Laws for Video Benchmarking

PhyGround organizes physics evaluation into 13 fundamental categories: (1) Gravity and vertical motion, (2) Collision dynamics, (3) Friction and surface interaction, (4) Fluid dynamics, (5) Buoyancy and displacement, (6) Elastic deformation, (7) Fragmentation, (8) Material phase transitions, (9) Magnetism, (10) Thermal effects, (11) Pressure and compression, (12) Rotational dynamics, (13) Conservation laws (momentum, energy, angular momentum).

Each category includes specific test scenarios that isolate the relevant physical principle from confounding factors. For gravity testing, we use single-object free-fall scenarios with fixed camera and uniform background to eliminate variables like collision, friction, or wind resistance.

This decomposition allows us to pinpoint which physical laws a model understands well and which require further development. A model might demonstrate strong performance on gravity and collision but fail at fluid dynamics, revealing specific capability gaps.

Each PhyGround category requires purpose-designed test videos. For gravity, prompts must eliminate confounding factors: uniform background, static camera, single object, no additional forces. For collision, you add complexity: multiple objects, varying mass, varying collision angle.

Quality control is critical. Benchmark videos must exhibit ground truth physical accuracy, meaning they were either rendered using accurate physics engines or validated by domain experts. A benchmark video showing implausible physics would invalidate all scores derived from it.

Automation potential varies by law. Gravity violations can be detected via optical flow analysis (detecting sudden velocity changes). Fluid dynamics violations require more sophisticated analysis or human assessment. A comprehensive benchmark includes both automated checks and expert human review.

PhyGround framework builds on established physics benchmarking approaches. Related work includes: evaluation metrics for video generation (LPIPS, FVD, Frame Consistency Score), physics simulation validation (comparing against Blender/Unity outputs), and visual reasoning benchmarks (CLEVRER, Physion).

This methodology aligns with OmniVeo's Logic Score approach but provides finer-grained decomposition. Users interested in implementation details should refer to the Logic Score methodology documentation and the Evaluation Challenges section of the Gemini Omni analysis.

Note: PhyGround is a proposed framework rather than an officially published benchmark at this stage. Specific test scenarios and scoring rubrics continue to evolve as more models demonstrate physics capabilities.