In recent months, a large number of human motion generation models based on diffusion structures have been published. Due to the inherent defect of human motion datasets, the capabilities of these models are constrained, only allowing for the generation of motion sequences with restricted lengths. This paper attempts to employ a zero-shot approach to enable several motion diffusion models to generate long sequences, proposing a self-cure module which enhances the model's generative capabilities without additional training. For situations involving long transitions, we utilize Large Language Models to assist in the generation process. Extensive experiments on arbitrary-long human motion generation tasks demonstrate that our method is useful and effective.