We introduce a large-scale synthetic dataset and a fine-grained alignment framework for composed person retrieval, and provide a manually annotated benchmark test set for objective evaluation.