Abstract : Bayesian approaches assuming Gaussian models for the image patches have recently achieved impressive denoising performance. First, fixed models were used for all patches and then, performance was significantly improved by using per-patch models. Unfortunately, the local estimation implied by such an approach is particularly unstable for most inverse problems beyond denoising. In this work, we advocate the use of a hyperprior to model image patches, in order to stabilize the estimation procedure. Two characteristics of the proposed restoration scheme are: first that it is adapted to diagonal degradation matrices, and in particular to missing data problems (e.g. inpainting of missing pixels or zooming), and second that it can deal with generic noise models, particularly suited to digital cameras. As such, the scheme is especially adapted to computational photography purposes. In order to illustrate this point, we provide in the experimental section an application to single image high dynamic range imaging showing the effectiveness of the proposed scheme.