Valdi: Value Diffusion World Models for MPC

World models are crucial for enabling Model Predictive Control (MPC), but they require dynamics predictions that are both fast enough for real-time online use and expressive enough to capture uncertain future states. Diffusion models naturally excel at modeling uncertainty and multimodality, yet their iterative inference process typically makes them too slow for low-latency latent planning in MPC. This research bridges this gap by introducing Value Diffusion World Models, or Valdi. Valdi integrates end-to-end online training specifically for MPC with a latent diffusion dynamics model. The key innovation lies in making diffusion models practical for control by optimizing them for speed. Preliminary experiments conducted in the CarRacing environment demonstrate promising results. Valdi, even when utilizing just a single diffusion step during both training and inference, achieved performance comparable to a deterministic MLP baseline. The study also highlights an important trade-off observed in this setup: balancing the model's ability to represent multiple possible future outcomes (predictive multimodality) against its direct control performance.