Physics in 2-Steps: Locking Motion Priors Before Visual Refinement Erases Them

Phase components of image-to-video diffusion model outputs degrade roughly 18% between step 2 and step 50 of denoising, while magnitude stays stable — explaining why early-step outputs show better physical consistency than fully denoised ones. PhaseLock is a training-free framework that extracts motion priors from 2-step inference and locks the Fourier phase throughout the full denoising trajectory. The approach requires no retraining and targets physically implausible motion in existing I2V models.