Physical Consistency in Video Generation
Overview
This field asks whether generated videos merely look convincing or depict motion and interactions that could happen in the physical world. Approaches include guiding video diffusion with 3D geometry, simulated trajectories, motion priors, and structured event or state reasoning, as well as refining outputs with physics-based rewards or verification. Benchmarks measure properties such as trajectories, collisions, object-state changes, geometric consistency, and conservation laws, often alongside human judgments. Strong appearance and smooth motion still do not guarantee accurate physics, especially for complex interactions, counterfactual changes, and long-horizon state evolution; improving physical fidelity without sacrificing visual quality and prompt alignment remains a central tension.
Latest papers
191 papers36 papers in the last four weeks, up 112% on the four weeks before. 0.2% of all new papers.