Do generative video models understand physical principles?
This work addresses a foundational problem in AI for researchers and developers by showing that visual realism in video generation does not equate to physical understanding, highlighting a key limitation in current models.
The authors tackled the question of whether generative video models understand physical principles by creating the Physics-IQ benchmark dataset, which requires deep understanding of physics like fluid dynamics and optics. They found that current models, including Sora and Runway, have severely limited physical understanding unrelated to visual realism, though some test cases can be solved, indicating potential for learning from observation.
AI video generation is undergoing a revolution, with quality and realism advancing rapidly. These advances have led to a passionate scientific debate: Do video models learn "world models" that discover laws of physics -- or, alternatively, are they merely sophisticated pixel predictors that achieve visual realism without understanding the physical principles of reality? We address this question by developing Physics-IQ, a comprehensive benchmark dataset that can only be solved by acquiring a deep understanding of various physical principles, like fluid dynamics, optics, solid mechanics, magnetism and thermodynamics. We find that across a range of current models (Sora, Runway, Pika, Lumiere, Stable Video Diffusion, and VideoPoet), physical understanding is severely limited, and unrelated to visual realism. At the same time, some test cases can already be successfully solved. This indicates that acquiring certain physical principles from observation alone may be possible, but significant challenges remain. While we expect rapid advances ahead, our work demonstrates that visual realism does not imply physical understanding. Our project page is at https://physics-iq.github.io; code at https://github.com/google-deepmind/physics-IQ-benchmark.