EnactPhysExecuting Evolving Physical Processes
in Video Diffusion

Anonymous Authors

GitHub Models Dataset

Generalization

Method

EnactPhys architecture: latent object states evolve under physical parameters and interact; Reader and Writer exchange information with video features.

Object State Modeling

Maintain a sequence of latent states for each object, built from object-specific video features.

Object State Evolution

Force, gravity, and mass guide temporal updates. Friction and restitution guide inter-object interactions.

Video–State Exchange

The Reader updates states from video features. The Writer sends the evolved states back to guide video denoising.

Parameter Sweeps

Joint Physical Control

Composed Physical Events

Baseline Comparisons

Friction μ

Open video ↗
Comparison details & figure

Rows: EnactPhys, Phantom, ControlNet, and Text-only. Columns: five parameter settings. Full 49-frame sequences at their recorded 16 fps.

Markers indicate observed positions at a shared time within each panel. Connecting lines guide comparisons across settings; they are not temporal trajectories. The force panel uses a shared enlarged crop. Force and mass include closely spaced responses.

Five-parameter, five-level comparison of EnactPhys and three baselines