Adobe releases LDR: First video world model to extrapolate beyond training distribution

_akhaliq · x · 2026-08-25

Adobe has published "Learning How the World Evolves," introducing the Latent Dynamics Reasoning (LDR) model.

Key Features:

Learns the underlying dynamics of how the world evolves rather than directly predicting future pixel appearances.
Maps past frames into structured latent states and rolls them forward using kinematic integration.
Claims to be the first video world model that learns underlying dynamics purely from pixels and extrapolates them beyond the training distribution.

Code, models, and datasets have been open-sourced on Hugging Face and GitHub.

More from Multimodal