hugging-apps/latent-dynamics-reasoning

Latent Dynamics Reasoning (LDR)

Extrapolative video world model that rolls out future video frames from three initial conditioning images, reasoning about latent physical dynamics.

Upload three frames of a physical scene, pick the task type (uniform, parabola, collision, looming, bouncing), and the model generates ~29 future frames of predicted motion.

Based on: Learning How the World Evolves: Extrapolative Video World Models via Latent Dynamics Reasoning

• —Model (Apache-2.0)
• —Code
• —Project Page