We present a new method exploiting intensity variations of an image by photoclinometry in order to produce a high-quality, large digital elevation model (DEM) from an initial DEM with a lower lateral resolution such as a photogrammetric initial product. The method integrates an intensity model of the image based on a novel radiative transfer scheme and a realistic bidirectional reflectance distribution function (BRDF) model of the surface. Two carefully crafted regularization terms are also introduced ensuring the smoothness of the solution and the consistency with the photogrammetric information at large scales. The regularized inversion of the model based on an efficient numerical optimization scheme allows the method to recover details missing or degraded in the photogrammetrically generated DEMs. In addition, for processing challenging scenes, a coarse-to-fine strategy is put forward. The refined DEMs are validated and carefully characterized by multi-scale analysis. The experiments are performed using images of Mars acquired by the Context Camera (CTX) and the High Resolution Imaging Science Experiment (HiRISE) of the Mars Reconnaissance Orbiter. The method insures a spatial resolution that is comparable to that of the imagery and allows measuring heights with a relative precision down to 15 cm with HiRISE imagery.