Component segmentation and determination of coke optical microstructure can effectively enhance the application of coke in industrial production and hence contribute to the achievement of China’s “dual carbon” goals. With the rapid development of artificial intelligence technology, current segmentation and recognition methods for coke optical microstructure primarily rely on deep learning-based semantic segmentation models. However,the segmentation accuracy of existing models is often limited by the quantity and quality of data samples, resulting in low performance. To solve this problem, this paper proposes a few-shot coke optical microstructure segmentation method using Artificial Intelligence Generated Content (AIGC). In the first step, an initial dataset of 12 base images is expanded into a dataset of 3 000 images using AIGC techniques. Secondly, to improve the segmentation accuracy of coke optical microstructures, Local Contrastive Learning is incorporated into the MoCov3 framework,and a Semantic Controller module is integrated into the backbone networks. This enhances the extraction of useful features and generates the MoCov3-CD semantic segmentation model. Finally, we conduct contrast and ablation experiments on the constructed dataset to evaluate and analyze the MoCov3-CD model. Experimental results show that using the dataset expanded through AIGC, the MoCov3-CD semantic segmentation model achieves a pixel accuracy (PA) of 88.03% and a mean Intersection over Union (mIoU) of 68.59%. Compared with other state-ofthe-art semantic segmentation models, our model achieves results closer to those of fully supervised segmentation models. Specifically, its PA and mIoU are 2.68% and 3.72% higher than those of self-supervised segmentation models on average, and 1.8% and 2.42% higher than those o nf semi-supervised segmentation models on average.