Image generation models are increasingly capable of producing realistic frames, but their ability to simulate physical dynamics over time remains unreliable. The new Chain-of-Time method addresses this by having the model generate a sequence of intermediate images during simulation, rather than predicting a distant future state in one step. This is inspired by human mental simulation and by in-context reasoning in large language models, and it requires no additional training—only a change in prompting at inference time.

The authors tested the approach on synthetic 2-D graphics and natural 3-D videos, covering rigid-body rolling, harmonic oscillation, and pulley systems. They found that Chain-of-Time stabilizes the physics simulated by state-of-the-art image generation models, producing more accurate long-term predictions than direct generation. The method appears to work by providing a form of working memory, letting the model chain together many short-term predictions into a coherent trajectory.

Beyond performance, the paper analyzes what the model actually represents at each time step. This reveals that image generation models can encode physical properties such as velocity, gravity, and oscillation behavior, even though these are not explicitly labeled. The authors note that, like humans, these models predict future world states well over short horizons, and the Chain-of-Time mechanism extends that capability into longer simulations. The findings suggest that prompting strategies can unlock physical reasoning abilities already latent in image generation models, offering a lightweight alternative to fine-tuning for simulation tasks.