Interestingly, we can even analyze the generated image sequences using traditional computer vision algorithms to see the specific trajectories models take during in-context physical simulation. Like chain-of-thought, these intermediate steps are often structured and interpretable.
