Patch-based methods are widely used in various topics of image processing, such as image restoration or image editing and synthesis. Patches capture local image geometry and structure and are much easier to model than whole images: in practice, patches are small enough to be represented by simple multivariate priors. An important question arising in all patch-based methods is the one of patch aggregation. For instance, in image restoration, restored patches are usually not compatible, in the sense that two overlapping restored patches do not necessarily yield the same values to their common pixels. A standard way to overcome this difficulty is to see the values provided by different patches at a given pixel as independent estimators of a true unknown value and to aggregate these estimators. This aggregation step usually boils down to a simple average, with uniform weights or with weights depending on the trust we have on these different estimators. In this paper, we propose a probabilistic framework aiming at a better understanding of this crucial and often neglected step. The key idea is to see the aggregation of two patches as a fusion between their models rather than a fusion of estimators. The proposed fusion operation is pretty intuitive and generalizes previous aggregation methods. It also yields a novel interpretation of the Expected Patch Log Likelihood (EPLL) proposed in [40].