The performance of model predictive control (MPC) relies heavily on precise system models and well-designed cost functions, which limits its applicability to complex nonlinear systems. To address this limitation, this paper proposes an intelligent MPC scheme that integrates diffusion models (DMs) with generative adversarial imitation learning (DGAIL). By using the strong distribution modeling capability of the DM, the proposed method learns an implicit reward function more accurately from expert demonstrations, thereby significantly reducing reliance on manual reward design and expert prior knowledge. It also combines the benefits of model-free learning and model-based optimization. Furthermore, a particle swarm optimization (PSO)-based online update strategy is introduced to efficiently solve the constrained nonlinear MPC problem. Experimental results across multiple benchmark simulation demonstrate that the proposed approach outperforms existing imitation learning and reinforcement learning methods in terms of control performance and safety, confirming its effectiveness and generalization ability.