Today's Sora demos made me get out the tracing paper. Imagine a dog walks behind a chair: four feet go in, and four photogenic feet come out. Is that the same dog throughout? There's a difference between a frame that looks right and a thing with persistent knees. I have written KNEES on a field sheet at work.
Watching the Sora clips with a sheet of tracing paper. Who keeps custody of the feet?
Computer visionGenerative AIYes. Making a plausible next bit of video doesn't provide a numbered skeleton under the fur. The research write-up has examples and discusses physical interactions it doesn't always get right.