Reproduce “Physics in 2-Steps: Locking Motion Priors Before Visual Refinement Erases Them” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Woojung Han, Seil Kang, Youngjun Jun, Min-Hung Chen, Fu-En Yang, Seong Jae Hwang

PhaseLock is a training-free framework that improves the physical consistency of videos generated by Image-to-Video (I2V) diffusion models. It achieves this by extracting physically consistent motion priors from few-step inference and guiding the full generation process to preserve these dynamics, leading to an average 6.2-point improvement on Physics-IQ scores with minimal computational overhead.

Key insights

  • Physically consistent motion priors are rapidly established within an extremely small number of denoising steps (e.g., 2 steps) by I2V models.
  • This initial physical consistency is often lost during the subsequent full visual refinement steps due to 'phase erosion' in the latent representations, where the phase spectrum (encoding spatio-temporal structure) degrades significantly.
  • Preserving these early-stage motion priors through spatial-domain latent delta guidance is an effective and computationally efficient strategy for maintaining physical consistency without direct, potentially unstable, frequency manipulation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.