H3 video model reportedly fixes the gymnast-proportion problem and supports 3D motion reconstruction

andrew_n_carr · x · 2026-08-04

The poster says H3 is the newest video model to hit the timeline and claims it largely solves the long-standing “gymnast” problem once the prompt format is right. They say earlier models broke human proportions, but H3 produced convincing motion for a parkour stunt where the performer fails, and that the output could be fed into Comic 4.2 to reconstruct the motion in 3D.

They frame it as useful for simulation, prediction, blocking, and other motion-analysis tasks.

Original post →

More from Embodied

Embodied channel →