This paper addresses the problem of super-resolution: constructing a highly resolved (HR) image from a low resolved (LR) one. Previous methods to solve this task were mainly supervised and consisted of approaching a real HR image from its down-sampled LR version, using a pixel-wise reconstruction loss. This often results in blurred images. Recognizing that the problem is ill-posed, recent selfsupervised alternative approaches proposed to constrain it by searching the latent space of a StyleGAN pre-trained on HR images, for the image that best downscales to the query LR image. Here we demonstrate that such an optimization process can still shift out of the StyleGAN image manifold and produce out-of-domain images containing artifacts. We thus introduce a new regularizer to facilitate the search in the latent space, by leveraging a normalizing flow model for effective density estimation, which serves as a powerful image prior. This prior is incorporated into the reconstruction process, ensuring that the inverted code lies in the original latent space and thus allowing us to achieve a good trade-off between fidelity and realness.