I've been analyzing the output from several prompts focused on human motion, and the results are consistently off. The movement often violates basic kinematics.
For example, a prompt for "a person walking across a room and picking up a book" results in:
* Unnatural limb articulation, like elbow joints bending in incorrect planes.
* A lack of weight shift during gait cycles. The foot contact seems to "slide."
* Hand-to-object interaction is especially problematic. Fingers phase through solid objects or make impossible adjustments mid-grasp.
It's not just subtle. It breaks the physical plausibility needed for anything beyond very short, atmospheric shots. The temporal coherence for articulated figures seems significantly worse than for fluid or environmental elements.
You're definitely not the only one, and you've nailed the specific issues. The limb articulation problems are a huge barrier for any kind of professional use case.
I've seen the same thing, especially with hand interactions. It almost feels like the model understands the *idea* of a grasp, but doesn't have the underlying 3D spatial reasoning to execute it without phasing or sliding. Even simple objects like a coffee cup can look impossible to hold.
This makes me wonder if we're hitting a fundamental limit with the current training data. Maybe temporal coherence for complex skeletons just needs a different approach than for, say, flowing water or clouds.