Entity Set Expansion (ESE) aims to expand entities belonging to the same semantic category given only a small set of seed entities. Traditional ESE methods are primarily text-based and struggle to address two key challenges in MOOC scenarios: (1) Multimodal Semantic Dependency; (2) Generalization Issues.To facilitate research in this area, we construct MOOC-MESE, the first Chinese multimodal ESE dataset for MOOCs, which contains over 45k aligned image-text pairs. The dataset provides rich multimodal semantic information for entity expansion research in educational scenarios. Furthermore, we propose GMRM (Generative Multimodal Retrieval Model), a multimodal ESE framework that reformulates the task as a multimodal generative retrieval problem. By achieving multimodal alignment and fully exploiting cross-modal semantics, GMRM effectively mitigates generalization issues. Extensive experiments and analyses on MOOC-MESE demonstrate both the high quality of the dataset and the effectiveness of GMRM. The code is publicly available at: https://github.com/constantinus-xi/mooc-mese.