DiLA: Disentangled Latent Action World Models

Researchers introduced DiLA, a self-supervised world model trained on unlabeled video.

DiLA disentangles content from structure to learn continuous, semantically organized latent actions.

This improves action abstraction without hurting generation quality, preserving high-fidelity video synthesis.

The approach could make video-based world models more useful for planning, simulation, and action transfer.