CtrlCache: Accelerating Interactive Video World Models with Control-Aware Caching

This project looks at a training‑free, control‑aware caching framework that speeds up interactive world models, especially for high‑fidelity dynamic process simulations and advanced process control (APC) visualization. Conventional interactive surrogate models and spatio‑temporal simulators have high latency when they generate autoregressive chunks and perform iterative denoising. To fix this, we adapt the CtrlCache method to use control input changes—like step changes in valve positions or feed flow rates—to schedule computation dynamically. By comparing the structure of nearby simulation frames under different control settings, the framework sorts states into initial, transition, turning, or steady‑state phases. Full denoising is kept for important control transitions, while unnecessary high‑frequency calculations are cached and reused

during steady‑state periods. This work gives a clear guide for adding action‑aware scheduling policies to deep‑learning‑based process simulators. We evaluate performance by measuring computational speedup, how well the generated state trajectories match the true structure, and the accuracy of the control response. The results show a practical way to achieve real‑time interactive operator training and faster model predictive control scenario testing.

Message: Hi FE, I need help with “CtrlCache: Accelerating Interactive Video World Models with Control-Aware Caching” in “Chemical Engineering”