As deep neural networks (DNNs) are hard to be trained due to gradient vanishing, intermediate supervision is typically used to help earlier layers to be better optimized. Such deeply supervised methods have proved to be beneficial to various tasks such as classification and pose estimation, but it is rarely used for image super-resolution (SR). This is because intermediate supervision needs a set of intermediate labels, but in SR, these labels are hard to be defined. Experiments show that identity labels across the whole network, which are used for classification, will cause inconsistence and harm the final performance. We argue that ‘mediately accurate’ labels, i.e.relatively soft labels, are more suitable for intermediate supervision on SR networks. But labels in SR networks are of either completely high resolution or completely low resolution. To address this problem, we propose what we call pushing and bounding loss, which forces the network to learn better features as it goes deeper. In this way, we do not need to explicitly give any ‘mediately accurate’ labels but all internal layers can also be directly supervised. Extensive experiments show that deep SR networks trained in this scheme will receive a stable gain without adding any extra modules.