This PhD project focuses on the development and theoretical and empirical analysis of diffusion-based generative models as world models for reinforcement learning. The goal is to design scalable diffusion architectures that can learn rich, temporally coherent representations of environments, and to integrate these models into planning, control, and policy learning frameworks. The project will investigate how diffusion models can be used for forward simulation, uncertainty-aware prediction, and representation learning, as well as how they interact with model-based and model-free reinforcement learning algorithms. The research will combine methodological contributions, theoretical insights, and large-scale experimental validation, with applications ranging from continuous control to sequential decision-making in complex, high-dimensional environments.