To overcome the limitation that missing modalities impose on multi-spectral vehicle re-identification in practical surveillance, we propose CycleFlow, a cycle-consistent flow-based generative framework that synthesizes features of absent modalities from available inputs. CycleFlow employs a reversible flow network with cycle-consistency constraints, ensuring identity-preserving feature generation while bridging the cross-modal representation gaps. The generated features are further fused with original modality embeddings through multi-modal Transformers, producing a unified representation for robust vehicle re-identification. Extensive experiments on three public multi-spectral datasets, RGBNT100, MSVR310, and WMVeID863, demonstrate that CycleFlow maintains competitive performance under full-modality conditions and significantly outperforms existing methods in scenarios with missing modalities. Ablation studies and qualitative analyses further highlight the effectiveness of generated features and cycle-consistent learning in enhancing cross-modal alignment and generalization.