Detecting metaphors poses a significant challenge in natural language processing (NLP) field. The methodology for metaphor detection has traditionally been feature-based semantic analysis, which treats every metaphor as data of equivalent complexity, ignoring the learning difficulty inherently in metaphors. To tackle this challenge, we introduce Curriculum Learning with Difficulty Division (CurLD) to help the model allocate more attention to difficult samples. We first distinguish the difficulty of the metaphors from both the general perspective of the model and the specifics of the metaphor detection task itself. Subsequently, we design cyclic curriculum schedule to progressively fine-tune GPT-2 from easy to difficult to achieve generative detection. Our method is assessed on the VUA, MOH-X and TroFi datasets, achieving results that are competitive with those of earlier approaches.