Structure as a Harness: Toward Physically Grounded World Models

How can we make generation a process we can inspect and steer as a scene takes shape? Drawing on my work on compositional 3D generation and controllable visual synthesis, I will show how explicit representations of objects, relations, and intermediate states can serve as a structural harness for generation, providing meaningful points of intervention during synthesis. Building on these foundations, I will discuss ongoing research toward physically grounded world models, exploring how geometric and physical inconsistencies in generated videos can be localized and turned into feedback for learning.