LTX-Video, CogVideoX and Pyramid Flow fall short of simulators on eight criteria

The authors of a systematic review on arXiv analyzed 200 papers on generative world models from 2018 to June 2026 and compared them with traditional simulators across eight capabilities: asset construction, physics engine, interaction, controllability, stability, state feedback, diversity, and evaluation metrics. The study From Generation to Simulation: How Far Are World Models from Being True Simulators? shows that world models already functionally replace simulators in interaction and controllability for individual scenarios, but lag behind in formal guarantees of physical laws, structured state feedback, and reproducible long-term evolution. The most notable gap is state feedback: only 6 of 163 works with implementations provide an interface for querying object states or physical parameters during execution.

What it means

The review directly concerns the video generators from our reference — LTX-Video, CogVideoX, and Pyramid Flow belong to the video generation route, which the authors highlight as one of three main technical paths to world models. According to the fact sheets of all three tools, none of them provides a state feedback interface — the very shortcoming the review calls the most neglected. LTX-Video handles the image-to-video task, CogVideoX and Pyramid Flow handle text-to-video; in all cases, the output is a video sequence, not structured data about object positions or physical scene parameters. This aligns with the authors’ conclusion: generative models do not yet provide runtime access to entity states, which traditional physics and game engines do. Long-term evolution stability is another criterion where world models lag. The fact sheets of our tools contain no data on reproducibility guarantees or generation robustness over long horizons: developers do not claim such capabilities for LTX-Video, CogVideoX, or Pyramid Flow. Evaluation on downstream tasks, which the review proposes as one of six development directions, is also not reflected in the fact sheets. The review provides a useful framework for selection: if you need a simulator with physics and feedback, the video generators from the AI Video section currently cover only part of the path — generating visually plausible sequences, but not simulation as such.