Why AI Videos Still Feel Wrong

TL;DR

AI video motion feels unnatural not from lack of data, but because training on cartoons and unrealistic media teaches conflicting physics.

Key Points

• 1.AI video generation excels at photorealism but fails at motion. Frames look visually impeccable, but movement breaks the illusion — researchers confirm this is the core unsolved problem, not image quality.
• 2.More compute and data alone don't fix motion. OpenAI's Sora scaling from 1x to 32x compute improves results but never fully solves physics-accurate movement, disproving the 'just scale it' assumption.
• 3.Cartoon training data actively corrupts physics learning. A new paper found that cartoons — where characters pause mid-air and bodies bounce like rubber — are among the worst training samples, teaching AI conflicting physical rules.
• 4.Cutting bad training data, not adding more, achieves a 74.1% win rate. By filtering junk influences and fine-tuning on high-quality physics examples, the improved model beat the baseline in 850 human judgments across 50 videos and 17 participants.
• 5.The team compressed 1 billion+ AI parameters down to 512 numbers using Johnson–Lindenstrauss projection. This technique, also used in Google's TurboQuant, preserves relative distances in high-dimensional data while making motion-source attribution computationally feasible.