Based on kernel ISOMAP and multi-class multi-manifold ISOMAP, supervised kernelized multi-class multimanifold ISOMAP is proposed specially for the facial recognition, which not only preserves the generalization property of kernel ISOMAP, but also works in a supervised manner and solves the problem that the number of neural network’s weights to be tuned increases exponentially with input dimension and easily occurs overfitting while using ISOMAP-C for face recognition. The experimental results on some face datasets show that the proposed method is effective and robust for the number of training sample.