Even then, the balloon clearly changes between scenes (and it’s still leaving heavily on slo-mo).

If you can’t nail consistency, you’re dead in the water. This js why I’m a bit skeptical that this is going to overhaul the entertainment industry any time soon.

This is a known problem in generative workflows for AI vids, but solvable. Midjourney recently introduced a feature that does this for stills, and controlnets available for the comfyui ecosystem also can partially solve this, albeit with some hassle. I'm pretty sure if not OpenAI themselves others will follow with their foundation models.