Abstract-In this paper, aimed at the extensively existing problem of slowness in mainstream image super-resolutions, an efficient approach is proposed for super-resolution based on the extreme learning machine (ELM) for single-hidden layer feedforward neural networks (SLFNs). Features and issues (e.g. parameter selections) in the application of ELM are discussed, on the basis of which a general framework for a variety of super-resolution problems is proposed, and corresponding experiments are conducted. It is shown in the results that the proposed approach can achieve relatively good quality and much faster speed compared to traditional reconstruction-based super-resolutions, therefore the effectiveness of this method is demonstrated.