Example-based methods have demonstrated their ability to perform well for Single Image Super-Resolution (SR). While very efficient when a single image formation model (non-blind) is assumed for the low-resolution (LR) observations, they fail when a LR image is not compliant with this model, producing noticeable artifacts on the final SR image. In this paper, we address blind SR (i.e. without explicit knowledge of the blurring kernel) using Convolutional Neural Networks and show that such models can handle different level of blur without any a priori knowledge of the actual kernel used to produce LR images. The reported results demonstrate that our approach outperforms state-of-the-art methods for the blind set-up, and is comparable with the non-blind approaches proposed in previous work.