Many computer vision problems can be formulated in a Bayesian framework with Markov Random Field (MRF) or Conditional Random Field (CRF) priors. Usually, the model assumes that a full Maximum A Posteriori (MAP) estimation will be performed for inference, which can be really slow in practice. In this paper, we argue that through appropriate training, a MRF/CRF model can be trained to perform very well on a suboptimal inference algorithm. The model is trained together with a fast inference algorithm through an optimization of a loss function on a training set containing pairs of input images and desired outputs. A validation set can be used in this approach to estimate the generalization performance of the trained system. We apply the proposed method to an image denoising application, training a Fields of Experts MRF together with a 1-4 iteration gradient descent inference algorithm. Experimental validation on unseen data shows that the proposed training approach obtains an improved benchmark performance as well as a 1000-3000 times speedup compared to the Fields of Experts MRF trained with contrastive divergence. Using the new approach, image denoising can be performed in real-time, at 8 fps on a single CPU for a 256 times 256 image sequence, with close to state-of-the-art accuracy.