Diffusion models have demonstrated significant performance in generative tasks and have attracted widespread attention. However, the high computational cost of the noise estimation network and the iterative generation process limit their widespread application. Existing caching techniques reduce this burden without additional training, yet error accumulation during generation degrades image quality. To address this issue, we propose a Kalman Filtering Correction method for diffusion model acceleration, which is termed as KFC. In particular, an estimation strategy is designed to mitigate cache-induced errors at each generation step without modifying the architecture of the full diffusion model. It is compatible with existing caching mechanisms and enhances efficiency. Experiments on various diffusion models and benchmark datasets validate its effectiveness.