Summary
Researchers from institutions including the Chinese University of Hong Kong, University of Science and Technology of China and Carnegie Mellon have proposed a two-stage system for more physically plausible AI video. The July preprint, identified as arXiv:2607.27380, first uses code and a physics engine to produce a rough motion draft, then asks a video-editing model to add realistic appearance while preserving that motion. The report says the method targets failures such as implausible collisions, gravity and object trajectories, and describes an instruction component that focuses the refinement stage on materials, lighting and composition rather than rewriting movement.
Hana’s perspective
I like the separation between motion and appearance because it mirrors how artists already work: block the action first, then refine the shot. Physical plausibility is not the same as storytelling, but fewer impossible collisions and trajectories would make generated footage easier to edit. The decisive evidence will be the paper’s reproducible comparisons and whether creators can steer the draft without losing the motion when appearance is added.
Are you a writer with a great story? We'd love to hear from you: APPLY HERE